Short-term and long-term adaptive changes in prosodic comprehension

Amelia Ostrow; Chigusa Kurumada

doi:10.1037/xlm0001621

. Author manuscript; available in PMC: 2026 Jul 15.

Published before final editing as: J Exp Psychol Learn Mem Cogn. 2026 Jul 13:10.1037/xlm0001621. doi: 10.1037/xlm0001621

Short-term and long-term adaptive changes in prosodic comprehension

Amelia Ostrow ¹, Chigusa Kurumada ¹

PMCID: PMC13367404 NIHMSID: NIHMS2158721 PMID: 42441610

Abstract

Speech prosody richly encodes both linguistic and social meaning. However, the prosody-meaning mapping varies across contexts and talkers, contributing to the lack of invariance problem in speech perception. This paper extends a recent line of work targeting the adaptivity of the human comprehension system, mapping variable input to linguistic meanings in a talker-sensitive manner. Specifically, we investigate the stability of adaptation: Once a listener adapts to the prosody of a particular talker, how long does the effect last? We conducted a large-scale (N = 424) adaptation experiment using a two-session paradigm in which listeners go through a pre-test, exposure, post-test sequence in the first session, in which exposure and feedback disambiguate an ambiguous prosodic contour as a question or a statement in a between-participants manner. They were then invited back on the second day after a delay of three, five, or seven days to repeat the same sequence. The results provide novel evidence that adaptation is both rapid and stable. Listeners showed robust adaptation within the first few trials of exposure in the first session, and they maintained some adaptive shift even after a delay of up to seven days. This finding extends the scientific knowledge about the longevity of adaptive shifts seen across levels of speech hierarchy. We discuss possible memory mechanisms that support the long-term maintenance of cross-talker adaptation.

1. Introduction

Speech prosody is an important channel of both linguistic and social meaning. Subtle modulations in pitch and timing can sometimes drastically change listeners’ word recognition, syntactic parses, and pragmatic comprehension (Cole, 2015; Cutler & Swinney, 1987; Ladd, 2008; Wagner & Watson, 2010). What has puzzled researchers for decades is the variability in the production of prosody. Even under the same or very similar meaning, different talkers produce different prosodic contours. At the same time, different meanings can be mapped onto very similar prosodic contours across talkers. For example, when your friend Sarah says, “It’s raining outside,” with a slightly rising final pitch, she could be asking you about the weather outside, or she might be telling you that it is raining using a style often referred to as “uptalk” (Calhoun & White, 2025; Warren, 2017). What is communicated through prosody thus depends in part on who said it, in addition to what was said and how.

The idea that spoken language comprehension is a talker-sensitive process is not new (Brown-Schmidt et al., 2015; Nygaard & Pisoni, 1998; Samuel & Kraljic, 2009). In fact, the last few decades of speech science research have revolved around questions of within- and between-talker variability and how perception copes with it (for comprehensive reviews, see Johnson & Sjerps, 2021; Kleinschmidt & Jaeger, 2015; Nusbaum & Magnuson, 1997; Quam & Creel, 2021). Human speech varies in myriad ways due to the physiological, social, and linguistic characteristics of talkers, to the extent that the same word is rarely produced with the same acoustic details twice, across talkers and contexts. The success of speech perception and comprehension ultimately relies on one or more mechanisms that reliably and robustly map such variable signals to latent form and meaning (Xie et al., 2023).

1.1. How do listeners navigate the prosodic variability?

A cursory look at the prosodic signal is enough to appreciate its cross-talker variations. The main determinant of the perceived pitch, the fundamental frequency (F0), is known to vary with the talker’s vocal tract length, which correlates with their body size (Hansen et al., 2015). Because of this, listeners are above chance at estimating the height (Pisanski et al., 2017) and gender (Lee et al., 2010) of the talker, even from just a few seconds of speech. In forensic science, pitch range and cadence are also associated with talker identity, as are other rhythmic and timing cues such as tempo and speech rate (Bartle & Dellwo, 2015; Jessen, 2008; Lavan et al., 2024; Podesva & Callier, 2015). Acoustic cues characterizing one’s prosodic productions thus form a vocal identifier of indexicality.

Variability in the prosodic signal has been at the center of theoretical debates in the literature (Arvaniti et al., 2024; Bolinger, 1986; Grice et al., 2017; Ladd & Morton, 1997, among others). Because of the substantial within- and between-talker variability, each phonological representation of intonational prosody, such as various pitch accent categories and boundary tones, subsumes a considerable amount of internal variability. It remains an open question to what extent and exactly which variations should be part of the phonological theory of intonation and which should be discarded as non-linguistic or incidental fluctuations (Gryllia et al., 2018).

The depth of this question is further magnified in the emerging body of large-scale production data (Cangemi & Grice, 2016; Chodroff & Cole, 2019; Dellwo et al., 2015; Steffman et al., 2024). Cangemi et al. (2015), for example, recorded three types of pitch accents produced by five native speakers of Standard German. Not only did the talkers differ in the exact shapes of a given pitch accent, but they also differed in the types and number of cues they used for a given accent, as well as the selectivity of these cues across accents (Grice et al., 2017). These differences are common across languages and are further exacerbated by social or regional dialects (Arvaniti & Garding, 2007; Clopper & Smiljanic, 2011). Parallel work has also found large individual differences in comprehension, adding to the fundamental theoretical puzzle of how listeners so reliably perceive and interpret prosodic contours as they do in everyday speech communication (Cangemi et al., 2015; Jun & Bishop, 2015; Kurumada & Roettger, 2022).

Two insights have been developed to address this puzzle. One is to take a bottom-up, data-driven approach to identify natural clustering of acoustic-phonetic cues produced under an intended intonational meaning (Arvaniti, 2019; Arvaniti et al., 2024; Steffman & Cole, 2024). Rather than pre-selecting a small number of specific categories and cues to analyze, as has been traditionally done, this approach can potentially integrate a large number of possible cues and their covariations. Using clustering algorithms such as a Functional Principal Components Analysis, it characterizes how natural confluence of cues may support a finite number of prosodic representations. For instance, Arvaniti (2019) applied this approach to three pitch accent types in Greek: H(high)*, L(low)+H*, and H*+L, in the ToBI annotation framework (* indicates an alignment between the tonal target and the stressed syllable). By characterizing these representations as partially overlapping (as opposed to discretely separable) distributions of phonetic cues, Arvaniti successfully captured the reliable separation of these pitch accents as phonological entities as well as the underlying phonetic variability within each accent (for a similar approach to information structure encoding in German, see Lorenzen & Baumann, 2024).

The second insight builds on the first and attempts to explain the processing of these distributional inputs. To illustrate, consider the “It’s raining outside” example above. Substantial phonetic overlap between intonational categories means that a given instance of input can often incur ambiguity between two categories (e.g., Low vs. High tone) as well as between contextually relevant meanings (statement vs. question). For listeners to successfully process a specific instance of input, knowing the general population- (or language-) level constellation of principal components is insufficient. An adequate theory of processing must explain how a listener can resolve the ambiguity to arrive at the contextual meaning intended by a specific talker, despite the natural language variability where the number, types, and exact distributions of cue clusterings can and often do vary across talkers (Arnhold & Kyröläinen, 2017; Huttenlauch et al., 2021; Severijnen et al., 2024b).

A key to navigating this variability is the notion of adaptivity in prosodic processing (Bosker, 2022). Continuing with the example of “It’s raining outside,” your recognition of Sarah’s utterance as a question (vs. statement) likely draws on two types of prior knowledge derived from previous experience with her: 1) the expected distributions of prosodic input if she intended to produce a question rather than a statement (i.e., how does the utterance sound if it is a question?); and 2) the contextual probability that she will ask a question rather than make a statement (Kurumada & Roettger, 2022; McQueen & Dilley, 2020). These two types of knowledge can be updated with additional exposure, simultaneously but independently. Imagine that Sarah is an uptalk user who produces many question and non-question utterances with a non-falling terminal pitch. One way for you to accurately recognize her intended prosodic categories can then be to adapt your knowledge about expected shapes of pitch contours for questions and statements produced by Sarah. Another way is to change your overall decision biases, e.g. you might categorize a larger proportion of her utterances as statements (or questions) than you would with another talker. With such adaptive shifts at multiple levels of speech perception and comprehension, initially ambiguous or confusing utterances become more reliably recognized as non-questions, improving your overall comprehension of Sarah’s utterances (for a review, see Xie et al., (2023)).

Several recent studies began to test this hypothesis (Bosker, 2021; Nakamura et al., 2021; Patel et al., 2011; Roettger & Franke, 2019; Roettger & Rimland, 2020; Severijnen et al., 2021, 2023, 2024a).

Most relevant to the current study, Xie et al., (2021) used a three-phase experiment consisting of a pre-test, an exposure, and a post-test. In the pre- and post-tests, participants heard items sampled from an 11-step continuum between a question (“It’s cooking?”) and a statement (“It’s cooking.”). Over 180 participants were assigned to one of the three exposure conditions after the initial pre-test. In one, utterances that are midway between a question and a statement are disambiguated as questions (question-biasing condition). In another, the same prosodically ambiguous utterances are disambiguated as statements (statement-biasing condition). Finally, there was a control condition where participants did not hear ambiguous utterances (no-shift). This exposure phase contained 30 utterances, 50% of which were ambiguous, lasting for four to six minutes.

In post-test, listeners responded to the identical set of items as pre-test from a question-statement continuum. Participants’ responses showed a clear between-group divergence in the directions consistent with the exposure, suggesting that they adapted their categorization of prosody after a brief exposure. This result resonated with that of another recent study in which Dutch listeners showed perceptual recalibration to acoustic cues to lexical stress (Bosker, 2022). After exposure to ambiguous instances of minimal pairs (CAnon- kaNON) with disambiguating feedback, they provided opposite responses to the physically identical, a priori ambiguous, tokens. It has also been demonstrated that listeners can track the use of prosodic cues for lexical stress (e.g., F0 vs. intensity) across multiple talkers. Incongruence in the learned mapping between prosody and talker can impact the accuracy and latency of perceptual decision-making (Severijnen et al., 2021; 2023). These findings highlight the exciting possibility that prosodic encoding remains plastic in adulthood. Listeners may be constantly calibrating their understanding based on previous experience with the same or similar talker, thereby navigating substantial variability in input from different talkers (Kurumada & Buxó-Lugo, 2024).

One major open question is whether and how long these adaptive changes in prosodic comprehension can persist. In a real-world context of speech perception, the speed of adaptation must be balanced with its persistence and stability. From a functional perspective, if the listener cannot maintain adaptive shifts, the process must start over each time for the same talker, even a familiar one. Such a constant (re)adaptation could be resource-intensive and therefore detrimental to perception. It would also hinder useful socio-indexical generalizations (e.g., learning “uptalk” as a style) as well. On the other hand, storing all adaptive shifts exhaustively and indefinitely would tax memory, especially for unfamiliar talkers whose speech we will not encounter again. Thus, the maintenance of adaptation can reveal an optimal balance between perceptual benefits and memory storage limits, as well as the amount of variability and stability that the human brain expects in its linguistic environment.

1.2. Stability of cross-talker adaptation

This line of reasoning was motivated by a parallel literature on segmental speech perception (Eisner & McQueen, 2006; Samuel & Dumay, 2021; Sharf & Ohde, 1981; Zhang & Samuel, 2014; Zheng & Samuel, 2023). A seminal paper by Eisner and McQueen (2006) systematically tested the stability of adapted changes in recognition. Participants’ categorization of a five-step continuum from “f” (read as /ɛf/) to “s” (read as /ɛs/) was tested immediately before and after they listened to a four-minute story designed to bias their interpretation of an ambiguous sound toward either /f/ or /s/, and participants responded to the same continuum 12 hours later. For half of the participants, the first session began at 9 a.m. and the second at 9 p.m., and the timing was reversed for the other half to see any effects of sleep on adaptation. They concluded that in younger adults, perceptual learning is stable over 12 hours, regardless of whether the test is administered after a day or after a night of sleep, which appears to have no effect.

Subsequent work has tested for retention of adaptation over a longer time. Using a cross-modal priming paradigm, Witteman et al. (2015) tested whether adaptation induced in younger adults via a foreign accent lasts with two delay conditions: one day and one week after exposure to globally-accented Dutch words (i.e., words spoken with a native accent, control condition, [ɪ] → [ɪ]) or Hebrew-accented Dutch words (experimental condition, [i] → [ɪ]). They concluded that the length of the delay between exposure and test did not alter the priming effect at all, suggesting that adaptation to foreign accents is stable from one day to one week after exposure. This finding was corroborated by Zheng and Samuel (2023), who investigated lexically driven adaptation to /θ/ and /s/ as well as /ʃ/ and /s/. Even though the adaptation to these two pairs is not the same, it lasted for over a day and could still be observed after seven.

Saltzman and Myers (2021) also tested younger adults twice, with the second session occurring 7 to 16 days after the first. Like Zheng and Samuel (2023) and many others, they used lexical decision tasks as training and non-word phoneme categorization (/s/-/ʃ/) as testing. They found that the speed and the magnitude of adaptation in the second session were reduced compared to those in the first session, suggesting that the information from the first session influenced responses in the second session. Additionally, even though they did not explicitly test for this, they discuss the possibility that individual differences in cognition likely contribute to how adaptation, perceptual learning, and their stabilities are realized (see also Heffner & Myers, 2021).

To our knowledge, there is only one study that has tested the stability of prosodic adaptation over the course of a day (but see Severijnen et al., (2024a) for stability of talker-specific learning of lexical stress after a shorter delay within a day). Luo et al. (2024) studied Standard Mandarin Chinese listeners’ adaptation to tones using word recognition tasks. To induce adaptation, syllables that are disambiguated as different lexical items when produced with tone 1 (high flat tone) or tone 2 (scooping rising tone), like qin, were paired with other words with no such ambiguity, such as mu (tone 1) or gu (tone 2). Lexical feedback from the non-ambiguous word served as adaptors and the ambiguous word would be interpreted differently through the adaptation. Test consisted of an alternative forced-choice task and occurred 12 hours after exposure, either during the day or overnight, as in Eisner and McQueen (2006). Although their control experiment confirmed that adaptation was observed in an immediate test, none was maintained after the interval, either wakeful or overnight.

This novel finding presents a possible, intriguing contrast between the segmental and suprasegmental (prosodic) levels, whose processing has been shown to involve overlapping but independent brain networks (Gnanateja et al., 2025). It is possible that their adaptation also involves different areas and networks as well. The stability of intonation adaptation is of particular interest because the key perceptual feature of intonation—vocal pitch—is perceived relative to each talker’s baseline (e.g., high vs. low) and mapped onto meaning primarily via its complex movements (e.g., rise vs. fall) and alignment with lexical level prosody rather than stable states. Moreover, such relatively broad-scale contour movements subsume all the variability at the lower levels, such as phonemes, words and phrases. It is currently unknown whether and how such dynamic prosodic representations can be stored in a talker-sensitive manner. To contribute to this body of work, we investigate American English intonation contours for questions vs. statements, which provide a useful combination between the dynamic pitch movements of intonational prosody and the relatively clear semantic distinctions accessible to naïve listeners.

1.3. The current study

The current study tests the retention of prosodic adaptation with up to seven days of delay between two sessions. In the past, studies on long-term effects of adaptation were hampered by logistical difficulties of subject recruitment, testing, and retention. We overcame the hurdles by conducting our study online (see also Heffner et al., 2022). We recruited 1360 subjects in total. As we describe below, we implemented multiple cautionary measures that ensure the experimental control and data quality. Participants were first given a brief headphone check, and only those with 100% accuracy proceeded to the rest of the experiment. After this screening, 424 participants were included in our analyses. An additional discrimination task also ensured stimuli audibility.

An added benefit of the online testing was to include participants with varying demographic backgrounds and socio-indexical features, more representative of the general population compared to an on-campus participant pool. This representativeness, however, also introduces uncertainty in the data analysis. In particular, participants could be differentially familiar with the type of intonation contours used in the study or the social variant of “uptalk”, which is typically associated with younger talkers from the west coast. This is critical for us because the familiarity may lead to biases in their categorization, which could have cascading effects on the adaptation and its retention. To this end, we have implemented a short questionnaire about the perceived regional dialect and age of the talker. This is, however, an inquiry secondary to the main question about the stability of adaptation. We will report on the descriptive statistics at the beginning of the result section and provide more detailed information in Appendix I.

An overview of each section of the experiment and its purpose is as follows. First, the headphone check screened out people who did not use headphones. The adaptation task, which included a pre-test phase, and exposure phase, and a post-test phase, came after this. During only the first session, the participants completed a questionnaire about their estimates about the speaker’s age and region of origin, which we used to see if participants’ assumptions about the speaker change their baseline interpretations of the adaptation stimuli. After the adaptation section, the participants completed a discrimination task. To robustly interpret the results of the adaptation task, we analyzed their performance in the discrimination task to test whether the six different between-participant groups (three delay conditions × two biasing conditions) differed in their basic perceptual sensitivity, and whether their perception of the stimuli’s low-level acoustics changed over the two sessions.

2. Methods

The current study was not preregistered, but the experimental and analysis methods were carefully validated by published work (Xie et al., 2021; Kurumada et al., 2024), and all parameters were planned prior to inspection of the data. The recruitment and testing procedure was approved by the institutional review board at the University of Rochester (Study ID: 00000764).

2.1. Participants

1360 individuals (aged 18-35 years old) were recruited via the online platform Prolific (https://www.prolific.com/). 541 (39.8%) successfully completed the headphone check (more details below) and completed the first session of the experiment. They were then randomly assigned to three between-participant conditions where the second session occurred three, five, or seven days later (i.e., three between-participant “delay” conditions). A total of 446 participants returned after these delays and completed the second session (90% after the three-day delay, with one participant not invited back because of technical difficulties; 77% after the five-day delay, with another participant not invited back because of technical difficulties; and 82% after the seven-day delay).

Of the 446 who returned, only those who successfully completed the second session were included in the final analysis (N = 424, Table 1). Each combination of delay and bias condition had more than 60 participants, which was necessary to achieve the statistical power computed based on Xie et al. (2021). Each participant received $4 per session. All participants were self-identified monolingual native speakers of American English, currently residing in one of the 50 U.S. states. See Appendix I for a breakdown of the participants’ self-reported genders.

Table 1.

Total number of participants in each session, by delay and bias. Bolded numbers indicate the participants who completed both sessions and were included in the final analysis.

Delay	Session	Question-biasing	Statement-biasing
3-day delay	Session 1	75	73
	Session 2	64	63
5-day delay	Session 1	80	76
	Session 2	67	69
7-day delay	Session 1	97	91
	Session 2	82	79
Sum	Session 1	252	240
	Session 2	213	211

Open in a new tab

2.2. Stimuli

2.2.1. Adaptation task stimuli

Sixty-six test and exposure stimuli were originally created by Xie et al. (2021). A male native speaker of American English naturally produced sentences of the form “It’s X-ing” (e.g., “It’s cooking”) and questions of the same form (e.g., “It’s cooking?”). He did this with six verbs: booting, cooling, cooking, losing, muting, and moving. The two tokens the speaker produced for each verb were synthesized in the phonetics analysis software Praat (Boersma & Weenink, 2012) into an 11-step continuum, ranging from prototypical statements (Step 1) to prototypical questions (Step 11) (Figure 1A). The continuum was created based on a natural recording of a canonical statement and question produced by a native speaker of American English in his 20s. We interpolated the duration and fundamental frequency (F0) of each syllable (“it’s,” “X,” and “ing”) and resynthesized the 11-step continuum using the pitch-synchronous overlap-and-add algorithm implemented in Praat (Boersma & Weenink, 2012).

Figure 1. — Experimental stimuli for the adaptation experiment. A. 11-step continua were created by duration and pitch of each syllable of “It’s X-ing” utterances. The plot is reproduced based on Figure 12 in Xie et al., (2021). B. Two between-participants bias conditions included ambiguous and unambiguous stimulus tokens, each followed by disambiguation feedback. The dots schematically represent the exposure stimuli given in each of the conditions, and the colors represent the feedback (blue = statement, red = question). Adaptive shifts of the categorization function in the pre- vs. post-tests are expected to occur in the directions corresponding to the continuum steps from which the ambiguous and unambiguous tokens are sampled, as well as the associated feedback.

The stimuli were normed by 57 native listeners of American English to validate that Step 6 elicits categorization judgments close to 50% question responses (i.e., ambiguous between statement and question, Xie et al., 2021, Appendix II). The two end point stimuli (Steps 1 and 11) as well as the most ambiguous stimuli were used as the adaptors during exposure, associated with disambiguating feedback (Figure 1B, for more detail, see 2.3 Procedure). If adaptation occurs, participants’ post-test responses should deviate from their pre-test responses in opposing directions between the two between-participant biasing conditions.

2.2.2. Discrimination task stimuli

Thirty-six stimuli for the discrimination task were originally created by Kurumada et al. (2024). They extracted the final syllable (i.e., “-ing”) from the 11-step continuum used in the test phase (i.e., “It’s cooking”) (Figure 1A). These 11 steps of -ing stimuli were combined into nine pairs of items that are two steps apart from each other (e.g., step1-step3, step2-step4). For each pair, we created 36 unique triplets such as 1-1-3 and 3-1-1 by crossing (a) nine stimuli pairs (e.g., 1-3, 2-4, 3-5), (b) two oddball identities (e.g., 1 or 3 for the 1-3 pair), and (c) two oddball locations (i.e., First or Last). Each subject listened to these 36 stimuli in a randomized order, and this procedure was repeated twice for a total of 72 trials. All stimuli from the adaptation and discrimination tasks are downloadable from OSF https://osf.io/3xvsh/?view_only=59f92f1227f44dbe848c4e3dd33f1ead (Ostrow & Kurumada, 2025).

2.3. Procedure

The experiment was implemented using the online task builder Gorilla.sc (https://app.gorilla.sc/). Each participant took part in the experimental tasks online in a self-paced manner. In session 1, participants first proceeded through an introduction, during which they took part in the six trials of headphone check task (Milne et al., 2021) (Figure 2). They listened to three bursts of white noise, one of which also includes the same white noise, phase shifted, presented in one ear. The task for listeners is to identify it in a three-alternative forced choice question. Only those who achieved the 100% response accuracy in the six trials proceeded to the experimental tasks. Note that this is a stricter criterion than that recommended by Milne et al. (2021). In fact, only 39.8% of participants met this criterion to proceed to the main part of the experiment (see 2.1 Participants). We adopted this strict threshold to ensure the stimulus audibility and task compliance. While the headphone check was implemented in both sessions, rejection occurred only in the first session to avoid further data loss in the current costly two-day procedure.

Figure 2. — Experimental procedure for session 1. After a delay of 3, 5, or 7 days, the same procedure was repeated a second time, except for the rejection based on headphone accuracy and the age/region questionnaire that followed the pre-test. The red and blue icons indicate the between-participant assignments to the question-biasing and statement-biasing conditions, respectively.

The adaptation task begins with the pre-test block. Participants heard a token of “it’s cooking” from the question-to-statement continuum and responded to a two-alternative forced choice judgment by selecting between two printed words on the screen (“Asking” and “Telling”). They heard each token four times (11 steps * 4 = 44 trials) in a randomized order with no feedback. After the pre-test, only in session 1, they were asked two questions regarding the talker’s (perceived) age and geographical accent. A map outlining the six regions (= Northeast, Midwest, South, West, Pacific (Alaska) and Pacific (Hawaii)) as well as an option “Not from the US” (Appendix I). Independent of these “inferred” gender and geographical region of the talker, we also collected the gender and current location of the participants (= listeners). The descriptive statistics are provided in Appendix I for interested readers. While a full combinatorial exploration of the (perceived) social variables of talker and listener can be instructive, it is beyond the scope of the current investigation. Critically, it is orthogonal to our main hypothesis regarding the maintenance of adaptation. We therefore do not discuss it further.

The 30 trials in the exposure block consisted of 15 ambiguous and 15 unambiguous item (five verbs repeated three times as ambiguous and unambiguous items). Unlike in a test block, stimuli were sampled from two points along the continuum (Figure 1B), and participants received explicit disambiguation feedback after each response (e.g., “That’s right. He was telling you something,” “Actually, he was asking you something”). In the question-biasing condition, items were sampled from Steps 1 and 6, associated with statement and question feedback respectively. In the statement-biasing condition, items were sampled from Steps 6 and 11, associated with statement and question feedback respectively. In other words, the prosodically identical items from Step 6 were disambiguated as questions and statements, alongside the clear instances of statements (in the question-biasing condition) and questions (in the statement-biasing condition). The test item “It’s cooking” did not appear in the exposure block.

Finally, participants completed the post-test block identical to the pre-test. They responded to another 44 trials of the two-alternative forced choice trials without feedback. In the following discrimination task, participants responded to 72 trials of oddball detection questions. In each trial, they heard a triplet of the “ing” syllables from the “it’s cooking” continuum and selected the oddball by clicking on one of the printed texts on the screen (“First” or “Last”). The left-right placement of the printed words was fixed throughout the 72 trials. Upon completion of these tasks, participants completed a demographic and language background final questionnaire.

Using Prolific’s “custom allow” function, we invited the participants back to session 2 after three, five, or seven days. The overall procedure used in session 2 was identical to session 1 with the age-dialect question and the final demographic and language background questionnaire removed. The average participation time was 23 minutes in session 1 and 21 minutes in session 2.

2.4. Predictions

Data from the discrimination and the adaptation tasks were analyzed separately, each with a set of mixed-effects logistic regression models (lme4 package) in the statistical language R (R Core Team, 2024). We will provide specific fixed effects and random effects for each model below. The discrimination task was designed as a control task to test whether participants’ sensitivity to subtle intonational variations differed between the two sessions as well as across the three delay conditions. Given the uniform sampling of stimuli and the absence of feedback, we predicted no significant effect of sessions (session 1 vs. session 2) or across the delay conditions.

For the main adaptation task, we tested the following three predictions.

Prediction 1: Listeners adapt their prosodic categorization within each session (i.e., one day)

Replicating the previous results (Kurumada et al., 2024; Kurumada & Buxó-Lugo, 2024; Xie et al., 2021), we expected that the brief training, lasting for only four minutes, would result in increase and decrease of the question responses in the question-biasing and in the statement-biasing conditions respectively. This was tested via a main effect of block (pre vs. post).

Prediction 2: Listeners maintain their prosodic adaptation for two sessions

Maximal maintenance of the adaptive shift would predict pre-test responses in session 2 to be identical to post-test responses in session 1. Partial maintenance would predict pre-test responses in session 2 to fall between pre- and post-test responses in session 1. Finally, no maintenance would predict pre-test responses in session 1 and session 2 to be equivalent. This was tested via a block×session interaction.

Prediction 3: The maintenance of the adaptation decays over time

If we observe maintenance of adaptation across sessions, we further investigate its possible decay function. If the magnitude of maintenance decreases across the three delay conditions (i.e., three, five, and seven days), then listeners’ pretest categorization judgments in session 2 will be most different from those in session 1 after a 3-day delay than after a 5-day and a 7-day delay (in that order). This was tested by a three way, block×session×delay interaction.

Because predicted effects of the two bias conditions (question- vs. statement-biasing) were independent of each other, we analyzed the two conditions separately.

3. Results

The results of the discrimination task are presented first as this served as a control task. We will then present the results of the social inference questionnaire. This asks if, prior to exposure, participants differed in their expectations for the talker’s age and regional origin. A significant difference, if any, might be reflected in their categorization (i.e., question vs. statement) judgments across the two sessions. Finally, we will present and discuss the categorization results in the adaptation task across two sessions to test the three predictions above.

3.1. Discrimination task

As shown in Figure 3, the overall accuracy was consistently above 80% for all conditions, demonstrating that the acoustic differences in the adaptation stimuli were easily discriminated for each participant. As predicted, the ranges of within-group individual differences were also comparable across the delay and bias conditions, and stable across sessions. The repeated testing and the experience in session 1 do not appear to have qualitatively changed participants’ response behavior in this task.

To test whether the discrimination performances indeed remained stable, we used a generalized linear mixed model with family “binomial” and the logit-link function to examine the main effect of bias (factorial, contrast-coded, statement-bias = −1 vs. question-bias = 1), delay (factorial, sliding-difference coded), session (factorial, contrast-coded, session 1 = −1 vs. session 2 = 1), and the interactions between them. Participant was included as a random effect. We used the maximal random effect structure justified by the data, i.e., by-subject intercepts and random slopes for delay as well as by-item intercepts and random slopes for the interaction between delay, bias, and session. A model summary and an additional analysis can be found in Appendix II.

As expected, session was not a significant predictor of the discrimination accuracy ( $\hat{β} = 0.088, z = 1.535$ ), meaning that listeners’ sensitivity to the subtle differences between contours did not change between session 1 and session 2. Although there was a marginal effect of delay condition between the 5-day vs. 7-day groups ( $\hat{β} = .29, z = 2.097, p < .04$ ), there was no interaction between this and session. This suggests that the three between-participant groups differed slightly in their baseline of discrimination accuracy, but their performance did not improve or decline within each. Taken together, the results support the assumption that experience in the first session did not significantly alter the perceptual sensitivity to prosodic contours, or that no change was retained over a delay. In other words, changes across sessions in the adaptation task, if any, can be reliably attributed to listeners’ categorization judgments, rather than to changes in the low-level perception of the stimuli.

3.2. Social inference questionnaire

To examine whether listeners inferred a distinct social profile for the talker, we examined their responses in terms of inferred (a) age and (b) region of residence. The responses were given immediately after the pre-test block, before the exposure phase in session 1. The inferences therefore reflect the first impression during the first few minutes of exposure. Note that the original endpoint stimuli were recorded by a male native speaker of American English in his mid-20s. He is originally from Long Island (New York) and had been living in upstate New York for four years at the time of recording.

Figure 4 summarizes the mean proportions of participants’ “question” responses by their responses in the questionnaire, i.e., inferred age and region of residence. Although the inferred age varied significantly across individual participants, there was no predictable relationship with their likelihood of answering “question” to items sampled from the 11-step continuum (R² < .01). Similarly, a great deal of participants inferred that the talker is from the west (n = 175, 41.3%), which is often associated with the prevalence of uptalk. However, those participants’ likelihood of answering “question” is not lower than others (one-tailed t-test, t = −1.11, p > .13). While the average proportion of “question” responses appears higher among participants who thought the speaker was from the south, this group is too small (n = 12, 2.8%) to draw a decisive conclusion (for more information and discussion, see Figure A2 in Appendix I).

In summary, while the inferred social properties of the talker varied in both age and region, neither biased participants’ responses between questions and statements. We can thus safely assume that listeners’ post-exposure categorization responses were mostly influenced by the within-experiment exposure.

3.3. Adaptation task

Figure 5A displays the results from every bias and delay condition. As expected, within each block, the proportion of question responses increased monotonically from the lower to the higher end of the continuum. Also as expected, between blocks in each session, listeners’ responses shifted in the opposing directions depending on the exposure input and feedback (prediction 1). The session 1 (left) panels highlight that participants shift their interpretations of the ambiguous range (Steps 5, 6, and 7; in the gray box) upwards in the question-biasing condition and downwards in the statement-biasing condition. These results replicate previous findings: Listeners adapt their prosodic categorization to characteristics in the input (Bosker, 2022; Kurumada et al., 2024; Kurumada & Buxó-Lugo, 2024; Xie et al., 2021).

Most importantly for our current question, the pre-test responses in session 2 deviate from those in session 1 in the respective directions, resulting in a reduced magnitude of the pre- vs. post-test block difference in session 2 as compared to session 1 (prediction 2). Figure 5B summarizes the trajectory of the mean proportion of question responses collapsed over the 11 steps in each block. The maintenance was particularly salient in the statement-biasing condition: The responses for the post-test in session 1 and the pre-test in session 2 were nearly identical in all three delay conditions. This suggests that, contra prediction 3, listeners’ categorization of the prosodic continuum remained shifted after the long delay duration.

This is further illustrated in Figure 6, which presents the differences of likelihoods of question responses (converted to empirical logit) between the two pre-tests. The empirical logit conversion controls for the non-linearity of the bounded scale of proportions. If there was no maintenance of adaptation at all, the two would be identical, resulting in a difference of 0. In other words, the area under the curves in Figure 6 represents the magnitudes of maintenance in different bias and delay conditions. Note that, if listeners had simply stored the overall decision ratio irrespective of the phonetic and phonological information (e.g., “I increased the {question / statement} response in the last session”), then their pre-test responses should shift “up” or “down” uniformly across the continuum. The non-uniform changes shown in Figure 6 instead suggest that listeners’ adaptation and its maintenance were conditioned on the prosody of the test stimuli.

In the analyses that follow, we used a generalized linear mixed-effects model to test predictions 1-3. As mentioned above, we did so for the question-biasing and the statement-biasing conditions separately. We discuss the model results for the question-biasing condition first.

3.3.1. Question-biasing condition

We constructed a generalized linear mixed effect model predicting listeners’ “question” responses (question = 1, statement = 0) with family “binomial” and the logit-link function. Four fixed effect variables and their interaction terms were considered: session (factorial, contrast coded, session 1 = −1 vs. session 2 = 1), block (factorial, contrast coded, pre-test = −1 vs. post-test = 1), delay (3-, 5-, and 7-day delay, sliding difference coded), and continuum (numerical variable of 1-11, centered). The sliding difference coding allows us to test the possible decay of maintenance with the increasing delay, while avoiding the unrealistic linearity assumption of a numerical coding (i.e., the effect of 5- vs. 3-day delay might not be equivalent to the effect of 7- vs. 5-day delay). The centering of the continuum steps ensures that its interaction with the other two fixed effects is assessed at the center of the continuum, where the adaptation effect is expected to be most prominent. The final model included by-participant random intercepts and slopes for block. A model summary is provided in Table 2.

Table 2.

Model summary for the question-biasing condition. Fixed effects estimates of the generalized linear mixed-effects are derived for delay, session, block, continuum, as well as their interactions. Rows highlighted in blue are of interest that are significant at the α = 0.05 level.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	0.446	0.06	7.383	1.54e-13	***
Delay 5-day (vs. 3-day)	−0.185	0.153	−1.212	0.226
Delay 7-day (vs. 5-day)	0.199	0.144	1.381	0.167
Session 2 (vs. 1)	0.182	0.06	3.024	0.002	**
Block Post (vs. Pre)	0.366	0.028	13.066	<2e-16	***
Continuum	1.283	0.015	87.609	<2e-16	***
Delay 5-day:Session 2	−0.094	0.153	−0.614	0.539
Delay 7-day:Session 2	0.028	0.144	0.194	0.846
Delay 5-day:Block Post	−0.097	0.07	−1.379	0.168
Delay 7-day:Block Post	0.114	0.066	1.727	0.084
Session 2:Block Post	−0.07	0.028	−2.517	0.012	*
Delay 5-day:Continuum	−0.098	0.034	−2.865	0.004	**
Delay 7-day:Continuum	0.103	0.032	3.221	0.001	***
Session 2:Continuum	−0.024	0.014	−1.765	0.078
Block Post:Continuum	−0.079	0.015	−5.419	5.98e-08	***
Delay 5-day:Session 2:Block Post	0.065	0.07	0.923	0.356
Delay 7-day:Session 2:Block Post	−0.004	0.066	−0.056	0.955
Delay 5-day:Session 2:Continuum	0.042	0.034	1.222	0.222
Delay 7-day:Session 2:Continuum	−0.016	0.032	−0.495	0.62
Delay 5-day:Block Post:Continuum	0.038	0.032	1.16	0.246
Delay 7-day:Block Post:Continuum	0.002	0.03	0.073	0.942
Session 2:Block Post:Continuum	0.019	0.013	1.447	0.148
Delay 5-day:Session 2:Block Post:Continuum	0.012	0.032	0.359	0.72
Delay 7-day:Session 2:Block Post:Continuum	0.004	0.03	0.119	0.905

Open in a new tab

As expected, the main effects of continuum ( $\hat{β} = 1.283, z = 87.608, p < 0.001$ ) and block ( $\hat{β} = 0.366, z = 13.066, p < 0.001$ ) were highly significant. Listeners provided more question responses for the higher end of the continuum, and the exposure increased an overall proportion of question responses. That is, listeners adapted their prosodic categorization to the input given during the exposure phase (prediction 1). Critically, the negative two-way interaction between session and block was significant ( $\hat{β} = - 0.07, z = - 2.517, p < 0.02$ ). This suggests that the amount of shift between the blocks was smaller in session 2 than in session 1. A follow-up simple effect analysis confirmed that this was primarily due to the increased question responses in the pre-test in session 2, demonstrating the maintenance of the adaptation over the delay (prediction 2). Finally, the lack of a significant three-way (delay ×session × block) interaction indicates that the magnitude of the retained adaptation did not change depending on the delay. In other words, increasing the delay does not lead to any significant change or decay in the maintenance of adaptation (prediction 3).

3.3.2. Statement-biasing condition

The model specification was identical to those for the question-biasing condition except that this model predicts the statement responses (statement = 1, question = 0). This coding of the dependent variable makes the directionality of the block effect constant between the question-biasing and statement-biasing conditions, which aids the interpretation of the data. The final model included by-participant random intercepts and slopes for block. A model summary is provided in Table 3.

Table 3.

Model summary for the statement-biasing condition. Fixed effects estimates of the generalized linear mixed-effects are derived for Delay, Session, Block, Continuum, as well as their interactions. Rows highlighted in blue are of interest that are significant at the α = 0.05 level.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	0.807	0.05	15.995	<2e-16	***
Delay 5-day (vs. 3-day)	−0.357	0.127	−2.805	0.005	**
Delay 7-day (vs. 5-day)	0.106	0.119	0.893	0.372
Session 2 (vs. 1)	0.233	0.05	4.639	3.50e-06	***
Block Post (vs. Pre)	0.189	0.034	5.6	2.14e-08	***
Continuum	−1.279	0.015	−86.132	<2e-16	***
Delay 5-day:Session 2	0.011	0.127	0.09	0.928
Delay 7-day:Session 2	0	0.119	0	1
Delay 5-day:Block Post	0.031	0.085	0.368	0.713
Delay 7-day:Block Post	−0.119	0.078	−1.516	0.129
Session 2:Block Post	−0.205	0.034	−6.105	1.03e-09	***
Delay 5-day:Continuum	0.253	0.036	7.104	1.21e-12	***
Delay 7-day:Continuum	−0.077	0.03	−2.567	0.01	**
Session 2:Continuum	0.008	0.014	0.588	0.556
Block Post:Continuum	−0.007	0.015	−0.487	0.626
Delay 5-day:Session 2:Block Post	0.108	0.085	1.267	0.205
Delay 7-day:Session 2:Block Post	0.048	0.078	0.609	0.543
Delay 5-day:Session 2:Continuum	0.138	0.036	3.889	1.01e-4	***
Delay 7-day:Session 2:Continuum	−0.002	0.03	−0.073	0.942
Delay 5-day:Block Post:Continuum	0.147	0.035	4.194	2.74e-05	***
Delay 7-day:Block Post:Continuum	−0.086	0.03	−2.925	0.003	**
Session 2:Block Post:Continuum	0.025	0.014	1.844	0.065
Delay 5-day:Session 2:Block Post:Continuum	−0.037	0.035	−1.057	0.29
Delay 7-day:Session 2:Block Post:Continuum	0.025	0.029	0.841	0.4

Open in a new tab

As in the question-biasing condition, the main effects of continuum ( $\hat{β} = - 1.279, z = - 86.132, p < 0.001$ ) and block ( $\hat{β} = 0.189, z = 0.034, p < 0.001$ ) were both highly significant. Listeners provided more question responses for the higher end of the continuum, and exposure and associated feedback increased the overall proportion of statement responses in all sessions (prediction 1). Similarly to the question-biasing condition, the main effect of session was significant ( $\hat{β} = 0.233, z = 4.639, p < 0.001$ ). This is likely due to the large number of statement responses given in session 2 in the pre-test. Crucially, the significant negative two-way interaction between block and session ( $\hat{β} = - 0.205, z = - 6.105, p < 0.001$ ) supports the interpretation that the magnitude of adaptive shift was smaller in session 2 than in session 1. A subsequent simple effects analysis confirmed that this was primarily due to the higher proportion of “statement” responses in the pre-test of session 2 as compared to session 1, i.e., maintenance of adaptation (prediction 2). Finally, as in the question-biasing condition, the lack of a significant three-way (delay ×session × block) interaction suggests that the magnitude of the retained adaptation did not change as a function of delay (prediction 3).

3.3.3. Question- and statement-biasing conditions comparison

On inspection, Figure 5 and Figure 6 suggest that the maintenance looks stronger in the statement-biasing condition compared to the question-biasing condition. To test if this is the case, we ran a combined model with the bias condition (factorial, contrast coded, statement-biasing = −1, question-biasing = 1). For ease of interpretation, we coded the dependent variable according to the bias conditions (i.e., for the question-biasing condition, question = 1 and statement = 0, but for the statement-biasing condition, statement = 1 and question = 0). The final model included by-participant random intercepts and slopes for block. See Table A4 in Appendix III for the full results.

The interaction in question, bias × session × block, is significant ( $\hat{β} = 0.069, z = 3.17, p < 0.003$ ). This indicates that the effect of maintenance (= a smaller difference between pre-test and post-test in session 2 than in session 1) was more significant in the statement-biasing condition than in the question-biasing condition. The higher-order (four-way) interactions with delay were not significant (5-day vs. 3-day delay: $\hat{β} = - 0.015, z = - 0.608$ ; 7-day vs. 5-day delay: $\hat{β} = 0.014, z = 0.664$ ). That is, the statement-biasing condition supported more robust maintenance of adaptation than the question-biasing condition across all three delay conditions.

3.4. Exposure phase responses

To delve deeper into the trajectory of adaptation within each session, we analyzed the participants’ responses during the exposure phase. Figure 7 plots the mean accuracy and response times in the exposure phase, collapsed across the three delay conditions. (We provide by-delay condition figures in Appendix IV.) Recall that listeners received two types of adaptor stimuli: ambiguous and unambiguous, each of which was followed by feedback that characterized them either as a question or a statement. For example, in the question-biasing condition, an unambiguous falling contour was played and followed by “statement” feedback and an ambiguous (flat) contour was played and followed by “question” feedback. The analogue was true in the statement-biasing condition; an unambiguous rising contour was played and followed by “question” feedback and the same ambiguous contour was played and followed by “statement” feedback.

Figure 7. — Percent correct responses for each trial in order in the exposure phase by bias, and session. The left column is the question-biasing condition, and the right is the statement-biasing. The top row is Session 1; the bottom is Session 2. Red points and lines indicate trials that should be considered questions, and blue points and lines indicate trials that should be considered statements. Therefore, in the question-biasing condition, the blue items are unambiguous and red items are ambiguous; in the statement-biasing condition, the blue items are ambiguous, and the red items are unambiguous. Response times were calculated from the onset of the stimulus.

Across conditions, from the outset, listeners were at ceiling with the unambiguous stimuli. In contrast, and by design, they were initially less accurate with the ambiguous stimuli. However, their responses improved to near ceiling accuracy as quickly as within the first 10 trials for each delay condition in both bias conditions. Participants’ response times also decreased dramatically within the first five to 10 trials, corroborating the rapidly diminishing categorization ambiguity.

We used mixed-effects regression models on the response accuracy data. As with test responses, we analyzed the data from the question- vs. statement-biasing conditions separately. Fixed effects included the three delay conditions (3-, 5-, and 7-day delay, contrast coded), session (1 vs. 2, contrast coded), stimulus type (ambiguous vs. unambiguous, contrast coded) and trial order (numerical, centered) and their interactions. We used the maximal random effect structure justified. Details of the model summary are provided in Appendix VI.

In both the question- and statement-biasing conditions, we found three key fixed effects to be significant: (1) The ambiguous stimuli were less accurately recognized compared to the unambiguous stimuli, as intended by the design (question-biasing: $\hat{β} = - 1.63, z = - 16.159, p < 2 e - 16,$ statement-biasing: $\hat{β} = - 0.719, z = - 6.974, p < 3.07 e - 12);$ (2) the higher the trial number, i.e., the more trials the participant had already done, the more accurate the responses (question-biasing: $\hat{β} = 0.04, z = 5.378, p < 9.57 e - 09,$ statement-biasing: $\hat{β} = 0.055, z = 7.072, p < 1.53 e - 12$ ); (3) the degree of improvement was greater for the ambiguous stimuli than for the unambiguous stimuli (question-biasing: $\hat{β} = 0.055, z = 11.898, p < 2 e - 16,$ statement-biasing: $\hat{β} = 0.042 z = 7.294, p < 3.48 e - 13) .$ Critically, none of the main effects or interactions involving delay conditions was significant, indicating that the overall pattern of the results did not differ even after the longest delay interval. Combined with the pre-/post-test responses, the data highlight the remarkable flexibility and stability of prosodic processing. Only a few instances of feedback on the ambiguous stimuli led to adaptation that persisted for up to seven days.

Of note, two more interactions involving sessions were significant only in the statement-biasing condition: (4) The ambiguous stimuli were responded to correctly more in the second session than in the first session ( $\hat{β} = .528, z = 5.364, p < 8.16 - 8);$ and (5) and such difference was larger in the earlier trials before the responses plateau again ( $\hat{β} = - 0.022, z = - 4.755, p < 1.98 e - 6) .$ This is exactly what we expect if adaptive shifts from session 1 affected listeners’ responses from the beginning of the session 2 exposure phase. We conducted a conjoined model of the both biasing conditions to confirm that these interactions were indeed significant only in the statement-biasing condition (Appendix VI). The asymmetry between the question- vs. statement-biasing conditions is consistent with the analysis of the test block responses reported above, although we did not anticipate this a priori and currently have no explanation for it. We will return to this issue in the General Discussion.

4. General discussion

To our knowledge, this is the largest-scale, two-session experiment testing the maintenance of adaptive changes in speech perception. Consistent with predictions 1 and 2, we found that listeners (1) adapted their categorizations rapidly after brief exposure in the initial session and (2) maintained some of these adaptations over two sessions, although the magnitude of maintenance differed between the two biasing conditions. Contrary to prediction 3, the three delay conditions showed equivalent levels of maintenance. Below, we first discuss the implications of this finding, followed by possible factors that may have contributed to the longevity of adaptation.

4.1. Longevity of intonation adaptation

Given the substantial variability across talkers, it is functionally advantageous to maintain adaptive shifts for individual speakers’ productions. Even then, the longevity of adaptive shifts over up to a week was rather unexpected. When participants returned to the second session after a 3-, 5-, or 7-day delay, their pre-test responses were significantly different from their pre-test responses in the first session. This means that only a few minutes of exposure can have a lasting impact on listeners’ prosodic categorization. The absence of change in accuracy on the discrimination task suggests that the maintained effects were expressed primarily in categorization judgments—the level at which feedback was provided—without allowing strong conclusions about the locus of adaptation. We note, however, that the current experiment does not, and is not designed to, determine whether listeners maintained the exact sound-category mapping, the global likelihoods of the two categories, or a combination of the two. We discuss these possibilities in Sections 4.2 and 4.3 below to inform future studies.

The current empirical results are noteworthy for several reasons. First, a previous study that examined the stability of prosodic adaptation showed no evidence of retention after 12 hours (Luo et al., 2024). This could be due to the exact nature of the representations tested (intonation contours vs. lexical tones), language (English vs. Mandarin Chinese), the participant groups, or some other aspect of manipulations (the types and dosages of exposure) (for a review, see Cummings et al., in press). One likely source of the discrepancy is the form and strength of feedback. Luo et al. (2024) used a lexical decision task, where listeners focus their attention to the word vs. non-word judgments, and the lexical identity serves as implicit teaching signal. In contrast, our task used more binary, explicit feedback. It is possible that the encoding of talker- (or task-) specific prosodic productions in our experiment was better registered and retained as a function of the strength of the feedback. In addition, the contrast we used was presented as binary (question vs. statement), whereas lexical tones are inherently four-way choices for native listeners of Mandarin. The relative simplicity of the categorization in our current experiment may have been a driver of the longevity of the adaptive changes.

Regardless of the exact origin, the current results showed some of the longest retention times for both segmental and suprasegmental perception. This is particularly surprising given the design of the current experiment. Importantly, while the data collection by Luo et al. (2024) (and many other previous studies) took place in a laboratory setting, ours took place online. Participants used their own devices and took part in the task as part of their familiar listening routines. (All participants had completed other tasks on the same experimental platform with the approval rates above 95%.) This means that the context of the experiment as an episodic memory cue is overall weaker for our participants compared to those in Luo et al. (2024), who physically visited the lab for the two sessions. Our participants retained and recalled a talker’s prosody despite listening to tens and hundreds of others in the same or very similar listening environment, i.e., online, on their own device, in a familiar physical space.

Also of note, since the seminal work of Eisner and McQueen (2006), most studies in this area have used a delayed test with no immediate test. That is, no additional exposure occurred prior to this delayed test. In contrast, the current experiment (as in Experiment 2 of Luo et al., 2024), included an immediate post-test to assess exposure-driven shifts in listeners’ categorization judgments. These test items were sampled uniformly across the 11-step continuum, a type of stimulus known to potentially dilute or undo adaptation (Liu & Jaeger, 2018; Zheng & Samuel, 2023). In other words, the uniform distribution of the test stimuli itself becomes a part of the input statistics, which can override the exposure information. And this was likely to be the case in the current experiment. Remarkably, even after this counter-effect, retention was detectable in this experiment across listeners and conditions.

The adaptive changes and their maintenance can have profound implications for theories of prosodic processing. A common approach to analyzing intonation contour is to annotate and categorize each token of input as an isolated event. To the extent that the context is considered, it is typically limited to word sequences in an utterance (e.g., predictability or word surprisal, Aylett & Turk, 2004; Malisz et al., 2018), discourse and information statuses such as new vs. given (Arnold & Watson, 2015; Baumann & Grice, 2006), or broader social or task considerations (Buxó-Lugo et al., 2018; Holliday, 2021). Current results, as part of the growing body of work on talker-specificity in prosodic comprehension (e.g., Bosker, 2021; Roettger & Franke, 2019; Roettger & Rimland, 2020; Severijnen et al., 2021, 2024a) suggest that the talker identity could also provide a global context, which fosters interdependence between processing episodes. A brief exposure to a particular talker can alter the processing with the same (or a similar) talker not only today, but also a week later at least when context is held constant. Conversely, the same physical sounds a listener hears today can be parsed differently depending on the amount and types of past experience they have accrued. If so, how can we build a theory of prosodic processing that can integrate such temporal and experience-based contingencies?

This question lies at the heart of modern phonological and phonetic theory, whose aim is to elucidate speech perception in the face of cross-talker or cross-context variability. In their 2023 review of research on language production, Goldrick and Cole (2023) pinpoint an emergence of hybrid exemplar models as one of the major advances in the field over the past decade. These models combine detailed memory traces of individual speech experiences with emergent symbolic structures, such as phonemes and syllables. As such, they have the potential to address the limitations of traditional generative models, which ignore experience and variability, and pure exemplar models, which struggle with abstraction and generalization. Hybrid exemplar models assume that speech production, including prosodic production, involves constantly activating and reactivating episodic memory traces at the sound and word levels. Following classic exemplar-based theories (e.g., Goldinger, 1998), productions adapt to match recently heard pronunciations, especially for low-frequency words. In turn, perception accommodates these variations by drawing on relevant past episodes. Although Goldrick and Cole (2023) do not explicitly discuss how the hybrid nature extends to perception, for the exemplar retrieval in production to work, the perception system must be able to store lexically-conditioned and talker-specific details in the input. Experience can thus continuously update the exemplar cloud, aiding the robust processing of speech with linguistic, social, and indexical variations.

The current data is the first to demonstrate just how rapid and stable prosodic adaptation can be, suggesting that even a tighter link between speech processing and memory encoding is needed. In particular, the hippocampus has been shown to reactivate older memories as new ones are formed (Anderson & Hulbert, 2021; Kuhl et al., 2010), which likely influences how context-rich prosodic meanings are formed and stored. More broadly, continuous reactivation of exemplars in online speech processing warrants further investigation. In this light, an important open question is not whether adaptation should be maintained, but rather how or at what level of speech comprehension it can be maintained and when it should be forgotten or pruned to allow for efficient memory encoding and retrieval. Below, in Section 4.2, we begin a discussion to highlight questions that have emerged from the current experiment.

A brief note on an unexpected data pattern is in order. As reported in Section 3.3, the magnitude of maintenance was larger in the statement-biased condition compared to the question-biased condition. We currently have two hypotheses regarding this asymmetry. The first is that the shift endorsed in the statement-biasing condition was akin to a sociolinguistic variant of uptalk. Perhaps the listeners simply labeled the talker as a user of uptalk, which helped compress and retrieve memory for the ambiguous input tokens. To address this, we conducted an additional, albeit preliminary, analysis using the social inference questionnaire responses to test whether participants’ initial inferences about the age and geographical origin of the talker might have affected their adaptation and its maintenance (Appendix VI). However, the most comprehensive model failed to converge due to the overly complex fixed effect structure. To further test this possibility, we need a more direct manipulation of the perceived social identity of the training and test talker(s).

Another possibility is that the enhanced maintenance in the statement-biasing condition was due to the syntactic structure of the stimulus utterances. In the current experiment, all tokens had the construction “It’s X-ing”, which has declarative syntax. In other words, “adaptation” in the statement-biasing condition can be achieved by ignoring a potentially ambiguous contour rise and relying on the interpretation given by the syntax. In the question-biasing condition, on the other hand, listeners had to interpret the ambiguous contour as part of the “question” category, which could potentially be more demanding on memory. In our future work, we plan to address this issue by using simpler, syntactically unmarked, items such as noun phrases (e.g., “Coffee!” (as placing an order) vs. “Coffee?” (as taking an order)).

In summary, we reported the novel evidence of robust maintenance of intonation adaptation up to seven days post-exposure. The ability to maintain the learned prosody-meaning mapping would facilitate talker-specific (or context-specific) categorization of intonation and thus comprehension resilient to signal variability. This raises a class of exciting questions about the mechanisms underlying the representation and processing of intonation. Chief among these is about the memory systems that support short-term adaptation to environmental variations in the input and the retention and generalization of the adaptation over time. Below we discuss our current hypotheses to be explored in future investigations.

4.2. Memory encoding of intonation in the moment and over time

One key question is whether, and to what extent, listeners consciously stored the learned shift of categorization functions in their declarative memory. Theories of the declarative vs. the procedural models state that these two types of memory encoding rely on distinct brain areas and networks, with different anatomical and physiological underpinnings and cognitive consequences (Ullman, 2004). It is possible that the maintenance of adaptation found here relied heavily on a declarative or explicit memory encoding strategy, as we often do when we learn new words in an instructional setting. If so, the current results might not generalize to more subconscious or implicit processes supporting our everyday spoken language processing. Admittedly, it is difficult to draw a clear line between declarative and procedural memory encoding strategies as they work in concert, and sound category learning is known to recruit both (in a manner modulated by task and individual differences, Arthur et al., 2021; Quam et al., 2018). For us, the importance of this distinction hinges on the question: Does the current finding of long-term retention scale to more natural, spontaneous language use, or is it an artifact of our experimental manipulation?

As we discussed above, the current manipulation employs relatively explicit feedback (e.g., “That’s right. He was telling you something.” / “Actually, he was asking you a question.”)—a type rarely available in real-world communication. The feedback was given at a more or less predictable timing after participants made a two-alternative forced choice response. In this setting, our adult participants likely approached the task as one of rule-based learning, which is usually linked to declarative memory. Additionally, the current intonation continuum shifted linearly along the dimension of F0, although syllable duration also covaries with it. The uni-dimensionality, combined with explicit feedback, is conducive to declarative memory strategies (Goudbeek et al., 2009; Maddox et al., 2013; Quam et al., 2018). In this light, it is not as surprising to find the fast learning, a hallmark of declarative learning.

In contrast, story-based exposure used by Eisner and McQueen (2006) and a lexical decision task used by Luo et al. (2024) rely on more implicit contextual feedback, which is arguably more common in our daily language use. The listener’s attention is primarily drawn to the recognition of a word, and feedback about the target sound category occurs implicitly through the lexical access. Such implicit learning of input characteristics is often associated with procedural memory, the result of which can be long-lasting, provided that learning instances are systematically repeated (Earle & Ullman, 2021). Perhaps the exposure task in Luo et al. (2024) and others was more strongly associated with procedural memory encoding, while the specific dose of input was not sufficient to support stable maintenance of adaptation. If exposure was repeated over time, listeners might begin to form multi-dimensional memory representations, as seen in many motor learning tasks, which could ultimately be longer-lasting than declarative memory.

This leaves us with a puzzle. If the current task likely supported explicit encoding of the prosody-meaning mapping, how can the adaptation be so long-lasting? Declarative knowledge can be automated and eventually stabilized, but typically through repetition and practice (DeKeyser, 2020; Earle & Ullman, 2021). In the context of our current task, however, this is an unlikely explanation given that participants received no additional exposure to the talker between the two sessions. Self-induced practice is also infeasible, given that participants only knew that they would return for another session, without knowing what would happen then. It is therefore unlikely, although not impossible, that they consciously memorized the exposure stimuli with the feedback and regularly recalled them during the delay period.

Potentially, the input and feedback during the first session were sufficient to form a basic memory representation that could be stabilized by sleep-induced memory consolidation (Earle & Qi, 2022; Fenn et al., 2003; Xie et al., 2018). However, additional explanation would be required because not all memory traces can persist after a seven-day delay, particularly given the hundreds or thousands of intervening utterances. In particular, the explanation must consider that the memory representations to be maintained could be highly heterogeneous across listeners. This is because our exposure stimuli were taken from five different continua created from five sets of basic stimuli with different final verbs. The “ambiguous” tokens heard during exposure were in fact heterogeneous in their segmental and phonetic details, and critically, not identical to any of the test tokens (Appendix VII). What’s more, the exposure items were sampled from only two points on the 11-step continuum (Step 1 or 11 = unambiguous, Step 6 = ambiguous, Figure 2), whereas the test stimuli spanned the entire 11 steps. This means that adaptation within the first session already involves (a) extrapolation of an underlying categorization function from the sparse exemplars and (b) generalization or transposition of its shift to the test continuum (with no feedback during the post-test). Maintaining this adaptation over time requires listeners to retain the outcome of their own exposure-posttest generalization.

Therefore, a useful discussion must address how listeners compress their episodic memory for speech in a manner generalizable and retrievable after a delay. Do listeners retain exact memory traces of the test stimuli combined with their own responses from the post-test in session 1? Such exhaustive topographic memory storage is central to the influential exemplar-based theory and predicts a range of talker-sensitive aspects of perception (Johnson, 1997; Pierrehumbert, 2001; for a discussion on speech production, see Goldrick & Cole, 2023). Or does adaptation entail some form of distributional representation, such as updating beliefs about the category means and variances for a particular speaker or group of speakers? (Clayards et al., 2008; Kleinschmidt & Jaeger, 2015; Theodore & Monto, 2019). Evidently, low-dimensional parameterization (e.g. adopting the normality assumption) aids efficient memory encoding. In theory, all that a listener needs to remember from session 1 can be reduced to a single value, e.g., the mean F0 of the final syllables heard during the exposure phase. This allows participants to “compute cues relative to expectations (C-CuRE)” when they encounter the same talker (McMurray & Jongman, 2011). However, the normality assumption has sometimes been criticized due to the complex, non-normal cue distributions found in production data (Carvalho et al., 2021). There is still much debate about whether prosodic categories can be represented as means and variances, and if so, which cues should be used (Arvaniti, 2019; Kurumada & Roettger, 2022).

Thus, the puzzle still remains. With the current data, we cannot say exactly which aspects of the prosodic information were stored in memory or how some of it could be retrieved after one week. To make progress, future studies could implement more regimented delay conditions (e.g. after 2, 6, 12, or 24 hours) with and without an immediate post-test, include tasks designed to disrupt specific memory representations, and administer independent memory consolidation tests to benchmark individual participants’ performances. We will return to these future directions in Section 4.4. First, however, we must consider an important caveat regarding the current results. That is, the stimuli and procedure of our experiment may have been particularly conducive to a heuristic or strategy that supports maintenance. If so, what did this heuristic entail, and what insights does it offer? And most significantly, what aspects of the current results generalize to adaptive changes in prosodic comprehension in a more ecological context?

4.3. Did participants use task-specific decision rules or strategies?

Reviewers noted two design features that may have made maintaining adaptation particularly easy. First, using a single, homogeneous test continuum (e.g. 11 steps of “It’s cooking”) could have helped to consolidate their adaptive shifts in categorization judgments. Second, the clear and explicit feedback provided during the exposure phase may have led to a simple decision rule or response strategy such as, “When I hear ambiguous intonation, I select a {statement / question}.” Relatedly, some of our pilot study participants observed that the ambiguous adaptor stimuli—prosodically midway between statements and questions—often carried distinctive communicative meanings. Some perceived them as figurative or sarcastic (e.g., “It’s cooking” meaning “Obviously it’s cooking; why ask?”), while others described them as emphatic. Because rising declaratives typically occur in contexts with specific pragmatic presuppositions (Gunlogson, 2008), these ambiguous contours may have invited unique interpretations neither purely question- nor statement-like, making them especially memorable. We therefore take seriously the possibility that some portion of the observed effects may reflect changes in response strategies or decision-level biases, rather than (or in addition to) changes in perceptual representations.

At the same time, prior work in segmental speech perception suggests that directing explicit attention to stimulus ambiguity does not necessarily enhance perceptual shifts in phoneme categorization. For example, directing attention to ambiguity does not increase perceptual shifts in phoneme categorization (McAuliffe & Babel, 2016), and in some cases, cueing attention toward the stimulus rather than higher levels of linguistic meaning has been found to reduce generalization of adaptation. Additionally, factors relevant to task engagement, attentional or memory load, or reward incentives do not majorly impact the magnitude of adaptation (Drouin & Theodore, 2019; Liu & Jaeger, 2019; Mechtenberg et al., 2025; Zhang & Samuel, 2014, but see Jesse & Kaplan, 2019). While caution is needed when drawing an analogy to prosodic processing, this body of work suggests that explicit feedback and knowledge of stimulus ambiguity alone are unlikely to account for the robust and sustained adaptation observed in the current study. Instead, the current results may be most comprehensively interpreted as reflecting a combination of processes, including potential changes in perceptual encoding, category-level expectations, and downstream decision-level strategies. Recognizing this plurality of mechanisms is essential for understanding how listeners adapt to prosodic variability in real-world communication.

To better situate our current findings within a broader scope of possible mechanistic accounts, we consider three types of mechanisms that allow participants’ responses to shift before and after exposure within a session, as well as how shifts can be maintained over a delay period. Following the computational framework of Adaptive Speech Perception (ASP) proposed by Xie et al. (2023), we focus on: (i) changes in auditory or perceptual encoding of prosodic cues, (ii) category-level adjustments in the mapping between prosodic cues and meaning-bearing categories, (iii) decision-level response biases. It should be noted that these options are neither exhaustive nor mutually exclusive; rather, they represent independent mechanisms. That is, one or potentially all of which underlie adaptive speech perception, including the long-term maintenance of talker-specific prosodic perception and comprehension.

One possibility is that exposure led to changes in the encoding of prosodic cues, such that the acoustic cues relevant for intonation perception (e.g., F0 height, F0 excursion, temporal properties) were transformed or normalized relative to a context- or a voice-specific baseline. Such adjustments are often characterized as domain-general, in the sense that they operate over acoustic dimensions independent of linguistic categories such as phonemes, syllables, or words (Cychosz & Newman, 2023; Lobanov, 1971; McMurray & Jongman, 2011; Nearey, 1989; Severijnen et al., 2025). Effects on speech perception—such as contrast enhancement or shifts in categorization—then emerge as a consequence of these early transformations (Apfelbaum & McMurray, 2015). This type of mechanism has been invoked to explain why exposure to nonlinguistic tones can alter the perception of vowels (Maslowski et al., 2019; Newman & Sawusch, 2009; Wade & Holt, 2005) and fricatives (Chodroff & Wilson, 2020). Converging evidence from fMRI and intracranial ECoG studies also shows talker-normalized encoding in auditory cortex (Li et al., 2021; Tang et al., 2017). For instance, neural responses to physically identical vowel formants shift systematically depending on the mean F1 of their surrounding acoustic context, consistent with talker-normalized encoding in auditory cortex (Sjerps et al., 2019). Applied to the present task, this account predicts that listeners adjust their encoding of prosodic input relative to the talker’s baseline—established through the observations of the pretest and exposure stimuli such that subsequent tokens are interpreted in relation to this inferred reference point.

The computational parsimony and low memory demands of this account make it appealing. Within the ASP framework, cue normalization can be implemented with a single free parameter—the rate of shifting the expected marginal (i.e., grand) cue mean—rather than estimating category-specific means and covariances, as required for category-level adjustments (discussed below). This simplicity can be key to supporting both the speed of adaptation and the robustness of generalization. The main challenge for this account, therefore, is not its feasibility but its insufficiency. By averaging over or ignoring linguistic contents of speech, marginal normalization can become imprecise and, in turn, less useful for guiding categorization. For example, in the current stimulus set, the test tokens contained a lax vowel (“It’s cooking”), whereas all exposure tokens contained tense vowels (e.g., “It’s losing”; Appendix VII, Figure A5). As a result, the endpoint question token along the test continuum had a relatively shorter vowel duration and a steeper F0 rise, whereas an “unambiguous” question token during exposure exhibited a longer duration and a shallower rise. If listeners were to transpose information from exposure to test while disregarding these lexical and segmental differences, their inferred F0 mean would systematically undershoot the actual test realizations. For cue normalization to exert a meaningful influence on categorization judgments under such conditions, it is therefore likely to be combined with, or conditioned on, higher-level linguistic representations—a possibility we turn to next.

The second possibility—category-level adjustments—is often implicitly or explicitly assumed under the general banner of perceptual learning (e.g., Bent & Baese-Berk, 2021; Luthra et al., 2021; Norris et al., 2003; Samuel & Kraljic, 2009) or dimension-based distributional learning (Idemaru & Holt, 2011, 2020; Liu & Holt, 2015). Within the ASP framework, this mechanism is formalized as a process in which listeners learn talker-specific, multi-dimensional cue–category mappings (e.g., how F0 and durational values are associated with question versus statement contours) and update those expectations in light of new input. While this class of models has been implemented for binary judgments along a single cue dimension (e.g., center of gravity for the /s/–/ʃ/ contrast (Cummings, 2025)), for two-dimensional cue spaces (e.g., F0 and VOT for stop voicing (Xie et al., 2023), F0 and syllable duration for question–statement distinctions (Xie et al., 2021)), they can also be extended to N-alternative forced-choice tasks such as vowel identification in a multidimensional vowel space (Persson & Jaeger, 2023). As the number of cues and categories increases, however, the computational complexity grows substantially, because both the means and covariance structures of each category must be estimated and updated independently. This added flexibility is nonetheless advantageous, as different categories often exhibit distinct distributional properties. For example, question utterances typically show greater variability in F0 than statements, across both utterances and talkers, making it useful to track their distributions separately in order to tune into a particular talker’s production (Xie et al., 2021).

The advantages and disadvantages of this mechanism are, in many respects, the mirror image of those associated with cue normalization. Learning category-specific distributions is computationally costly and memory intensive, and adaptation may proceed slowly for categories that are relatively infrequent in the input. For instance, talkers and talker groups can vary substantially in how they produce a pragmatically informative pitch accent type (e.g., indicating uncertainty, incredulity or contextual contrast, Pierrehumbert & Hirschberg, 1990; Ward & Hirschberg, 1985). However, these pitch accents (by definition) are licensed by specific discourse and communicative contexts. To learn the exact ways and cue distributions, listeners might need a substantial amount of exposure to a given talker’s speech across than when they do so for a specific segment such as /s/, which may appear multiple times within a single word or utterance. Once such learning is achieved, however, category-level representations can yield highly accurate predictions of a talker’s productions for specific linguistic category. In the present task, the high frequency of exposure to the question–statement contrast, combined with explicit disambiguating feedback, makes it plausible that listeners were able to estimate category means and covariance structures during the experiment. If so, category-level learning—potentially in combination with lower-level cue normalization—could provide a straightforward account of both the observed shifts in categorization judgments and the high-fidelity maintenance of talker-specific realizations over time.

A third possibility is that adaptation primarily reflects decision-level response biases that operate downstream from perceptual encoding and category representations. Within the ASP framework, this mechanism is formalized as adjustments to post-perceptual decision parameters—such as category priors or response biases—that shape how perceptual evidence is translated into categorical judgments. These biases can reflect top-down or contextual constraints that influence listeners’ responses independently of the acoustic input itself. For example, in naturalistic settings, statements are far more frequent than questions. In situations of ambiguity or signal degradation where the perceptual evidence is not decisive, listeners may therefore rely on such category-level expectations to guide their inferences. In a controlled experimental setting, on the other hand, response distributions are often balanced by design, a meta-awareness of this balance can be a source of response changes (e.g., as “response equilibration” in Vroomen and Baart (2009), for a related discussion, see also Tamminga et al., (2020)). Contextually aware listeners are likely to rely on various category-level expectations to guide their response behaviors, and these expectations themselves can adapt as a function of experience.

This hybrid perspective—in which behavioral adaptation reflects contributions from both perceptual and post-perceptual mechanisms—is supported by neurobiological evidence linking changes in behavior to distinct loci of neural activity. For example, Myers and Mesite (2014) showed that early adaptive changes in responses along an /s/–/ʃ/ continuum were associated primarily with activity in downstream, decision-related regions (notably in right frontal and middle temporal cortex), rather than with immediate retuning of auditory–phonetic representations. With continued exposure, however, sensitivity to the shifted phonetic boundary gradually emerged in left superior temporal regions. This pattern suggests that post-perceptual decision changes can precede, and potentially scaffold, more signal-dependent changes in perceptual encoding.

As with cue normalization, the primary appeal of decision-level bias lies in its computational and representational parsimony. Adjusting decision criteria requires no storage of detailed auditory or acoustic traces and can therefore be implemented rapidly. At the same time, decision-level shifts alone are often insufficient to account for more nuanced patterns of adaptation (Clarke-Davidson et al., 2008) such as changes in category variance (Clayards et al., 2008; Theodore & Monto, 2019) or cross-category generalization (e.g., when exposure to one vowel influences the categorization of other vowels produced by the same talker (Weatherholtz, 2015)). The present results also rule out the most extreme version of this account—namely, that listeners ignored the acoustic input altogether and altered response biases alone. Such a strategy would predict uniform increases or decreases in question or statement responses across the entire continuum, including the endpoints. In contrast, the observed shifts in intercepts, coupled with unchanged responses to the endpoints (Steps 1 and 11), are incompatible with this explanation. When combined with residual sensitivity to acoustic cues, however, decision-level bias can support simple and effective heuristics (e.g., “When the intonation is ambiguous, respond {statement/question} according to the exposure condition”), which are easy to acquire and may be readily reinstated after a long delay. Thus, decision-level biases are best understood as contributing to adaptation without fully supplanting perceptual or representational mechanisms.

Disentangling the relative contributions and trade-offs among these multiple mechanisms is a central goal of our future work. It is particularly important to understand how each mechanism explains changes in the intercept and slope of a categorization function in response to specific combinations of prior expectations, perceptual features of the input, response format, and corresponding response noise levels (Apfelbaum et al., 2022; Xie et al., 2023). Furthermore, in light of the discussion about memory and encoding strategies (Section 4.2), it is also instructive to explore how listeners recognized and approached the task of understanding initially difficult utterances.

To obtain preliminary and necessarily subjective insight into what listeners may have retained from the initial session, we replicated the statement-biasing, 3-day delay condition with a new group of 56 participants and administered a post-experiment questionnaire to ask about their experiences including conscious response strategies. We selected this condition because it demonstrated robust maintenance in the original experiment, making it an ideal testing ground for eliciting introspective reports about participants’ response patterns. At the end of Session 2, participants were asked the following four questions:

“Overall, how easy or difficult did you find the task of identifying whether the speaker was ‘asking’ or ‘telling’?” (Followed by an optional free response text box.)
“Did you notice anything unusual or surprising about the speech stimuli during the ‘Asking or Telling’ part of the experiment?”. To those who answered “yes”, an additional question was presented (“If yes, was it related to the way the speaker used pitch or intonation?”)
“What did you remember from the first session (on Tuesday) when you came back to the study today? Please choose all that apply.”
When you came back to the “Asking or Telling?” task today, did you use any particular strategy to decide whether the speaker was “asking” or “telling”?”

This new group of participants provided the same robust adaptive shifts of categorization responses. A comprehensive and fully reproducible analysis of these questionnaire data is provided in our repository (https://osf.io/3xvsh/?view_only=59f92f1227f44dbe848c4e3dd33f1ead). In response to the final question regarding strategy use, 14% (8/56) of participants reported that they had developed and used their own rule or strategy, whereas 66% (37/56) reported that they mostly relied on their “gut feeling”; the remaining participants indicated uncertainty (7%, 4/56) or variability (13%, 7/56) in their approach. Free-response comments suggest that many participants attended closely to intonation, and some clearly articulated internally consistent response rules. However, most did not specify an explicit strategy, instead offering descriptions such as, “I tried to guess and think of myself saying it.” Taken together, these responses suggest two tentative conclusions. First, at least some participants likely adopted explicit or semi-explicit response strategies, which may correspond to shifts in decision-level category priors (e.g., increasing the likelihood of “statement” responses). Second, a substantial proportion of participants did not report any verbalizable rule or heuristic and appeared to engage with the task by responding to the perceived properties of the stimuli. Importantly, however, self-reported strategy use should not be interpreted as a direct or reliable index of the perceptual or decision-level mechanisms engaged during the task. Overall, these additional data underscore the heterogeneity of adaptive mechanisms both within and across listeners, reinforcing the view that prosodic adaptation is likely supported by multiple interacting processes rather than a single dominant mechanism.

When combined effectively, these perceptual and post-perceptual mechanisms may together provide an efficient means to maintain adaptive shifts over time. In the introduction, we noted that interpreting Sarah’s “It’s raining outside” may draw on the listener’s knowledge of (1) how Sarah uses prosodic cues and (2) how likely she is to ask a question vs. make a statement. The latter reflects a category-level decision bias—or a category prior in terms of a Bayesian inference framework, which can influence the outcome and confidence of perceptual decisions, especially when the input is ambiguous. Some existing formal models of adaptive speech perception (e.g., Hitczenko & Feldman, 2022; Kleinschmidt & Jaeger, 2015; Kronrod et al., 2016; Norris & McQueen, 2008) predict that adaptation can arise from updates to such priors, independent of changes to cue-category mappings. Computational simulations based on the ASP framework indeed support this view (Xie et al., 2023, Section 4).

This does not rule out, or preclude the need for, perceptual-level adaptation. Rather, it suggests that adaptation operates across multiple levels of the speech-processing hierarchy, with listeners adjusting both perceptual encoding and post-perceptual decision processes to optimize comprehension under memory and data constraints. When viewed this way, the formation of a rule or bias at the category level by participants may not be a superficial hack or cognitive shortcut. Rather, it can be a computationally efficient strategy that enables the retention and generalization of adaptive shifts without the need for exhaustive storage of examples. If so, this strategy could be an integral component of the hybrid mechanisms that support robust speech perception.

Understanding the nature of these multiple adaptive mechanisms is an open question left for future studies. Longevity, as well as rapidity, of adaptation will help identify cognitive, neural, and learning mechanisms underlying speech adaptation. A leading example of such efforts is work by Samuel and colleagues (Samuel & Dumay, 2021; Zhang & Samuel, 2014; Zheng & Samuel, 2023), who investigated different decay trajectories for perceptual recalibration and selective adaptation of multiple phoneme contrasts on multiple time scales (e.g., a delayed test after 25min, 5.5h, 12h, up to seven days). Overall, they found that effects of perceptual recalibration, albeit weak, can sometimes be found week later while selective adaptation can persist only for about 5.5h. They argued that these two types of empirical signatures of adaptive speech perception must rely on different perceptual and cognitive mechanisms, contrary to an earlier computational account that unifies both under rational expectation adaptation (Kleinschmidt & Jaeger, 2015). It is currently difficult to reconcile these views because the evidence comes from experiments with different sets of stimuli and other details such as the timing, number, and duration of inputs, feedback, and other experimental details. However, the insight of using adaptation decay as a window into its underlying mechanisms is promising and deserves further investigation.

In another parallel line of research on speech production, Kaschak et al. (2011) demonstrated that syntactic priming for a double object vs. a prepositional object construction (e.g., The teacher sent the student a message vs. The teacher sent a message to the student) can persist for a week. This finding has helped to clarify and endorse the role of error-driven learning in priming. Importantly, such a long-lasting priming effect could not be readily explained by a competing and influential account of residual or transient activation, which spans a much shorter time window (e.g., Pickering & Branigan, 1998). We expect that the current finding of long-term maintenance of prosodic adaptation may similarly provide a useful window into what changes as listeners accumulate speech input, and how this knowledge is stored and used in subsequent communication.

4.4. Limitations of the current study and future directions

Several aspects of the current experiments limit the interpretation of data. First, as discussed above, the feedback in the current experiment was highly explicit and thus limited in its ecological validity. Likewise, the homogeneity of the test sentences “It’s cooking” could have led to meta-reasoning about the task objective and explicit response strategies. To overcome this limitation, our future experiments will use the types of cues available in conversation. For example, in a dialog, an interlocutor’s response can reveal whether a preceding, prosodically ambiguous, utterance was interpreted as a question or a statement. In response to “We’re going !/?”, a response of “I don’t know.” is congruent if the utterance was understood as a question. In contrast, “OK.” is congruent if it was understood as a statement. In addition, recent studies have identified facial gestures that systematically bias listeners’ interpretation of utterances as questions vs. statements (Nota et al., 2023). Future studies should explore those discourse and contextual cues as well as co-speech gestures to simulate more naturalistic feedback that combine to guide intonation adaptation.

Second, the current online experiment lacked precise control over the timing of testing and thus cannot assess the effect of arousal at the moment of testing, as well as the specific role of sleep consolidation. Previous research on native and non-native sound category learning and adaptation suggests that sleep plays a key role in localized changes in synaptic strength and systemic changes in the organization of information (Xie et al., 2016). Thus, the exact nature and duration of encoding-decoding interval will be particularly important as we move beyond adaptation to a single speaker to generalization across multiple speakers. A tightly regimented testing schedule, such as that used in Eisner and McQueen (2006) and many subsequent studies, will prove useful for investigating intonation adaptation.

Finally, and relatedly, our current experiment used a single talker, confounding “talker sensitivity” with “context sensitivity.” It is unclear whether participants learned the features as unique to the exposed talker, or whether the adaptive shifts could also apply to other talkers. Recent studies (Bradlow et al., 2023; Cummings & Theodore, 2022; Kurumada & Buxó-Lugo, 2024; Severijnen et al., 2023) as well as classic studies have used more than one talker to address this question. In the future, we could further elucidate the role of attention and memory mechanisms involved in adaptive speech perception by manipulating the number and type of talkers in immediate and delayed tests. In addition, retention might be facilitated if the expected rate of encounters with the same talker is high, and hence “worthy” of memory resources. If so, input provided in repeated encounters with the same (or a similar) talker within a short period of time could lead to consolidation of long-term memory compared to the same amount of input provided sparsely over a longer period of time. Future tests of effective adaptation should take into account how people efficiently store, consolidate, and prune the signal-talker mapping created at the moment of processing.

5. Appendix

Appendix I. Participant Demographics

Table A1.

Participants’ genders by bias and delay conditions.

Gender	Bias	3-day delay	5-day delay	7-day delay
Agender	Question-biasing	1	0	1
	Statement-biasing	0	0	0
Female	Question-biasing	27	32	39
	Statement-biasing	31	33	44
Fluid	Question-biasing	0	1	0
	Statement-biasing	0	0	0
Male	Question-biasing	32	32	37
	Statement-biasing	29	34	33
Non-binary	Question-biasing	2	2	4
	Statement-biasing	3	1	2
Other (please specify)	Question-biasing	1	0	0
	Statement-biasing	0	0	0
Prefer not to answer	Question-biasing	1	0	1
	Statement-biasing	0	1	0
Sum	Question-biasing	64	67	82
	Statement-biasing	63	69	79

Open in a new tab

Table A2.

Participants’ current regions by bias and delay conditions.

Region	Bias	3-day delay	5-day delay	7-day delay
West	Question-biasing	11	11	13
	Statement-biasing	9	8	17
Northeast	Question-biasing	16	18	20
	Statement-biasing	15	18	22
Midwest	Question-biasing	14	11	17
	Statement-biasing	14	12	9
South	Question-biasing	23	26	32
	Statement-biasing	25	31	31
Pacific (Hawaii)	Question-biasing	0	1	0
	Statement-biasing	0	0	0
Sum	Question-biasing	64	67	82
	Statement-biasing	63	69	79

Open in a new tab

Figure A1. — We asked participants to identify the talker’s regional accent in a 7-alternative forced choice question

Figure A2. — Matrices of participants’ estimation of where the talker is from by where the participant is from. A. Proportion of how many people are from each region by where those participants thought the talker was from. B. Question rate for each combination of participants’ estimates of where the talker is from and where the participant is from.

Appendix II. Mixed effect logistic regression models for the discrimination task

Table A3.

Model summary for the discrimination task. Fixed effects estimates of the generalized linear mixed-effects are derived for bias, delay, and session, as well as their interactions.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	2.417	0.068	35.443	<2e-16	***
Delay 5-day vs. 3-day	−0.236	0.144	−1.635	0.102
Delay 7-day vs. 5-day	0.29	0.138	2.097	0.036	*
Bias Question vs. Statement	0.047	0.056	0.845	0.398
Session 2 vs. 1	0.088	0.057	1.535	0.125
Delay 5-day:Bias Question	−0.202	0.144	−1.402	0.161
Delay 7-day:Bias Question	0.16	0.134	1.195	0.232
Delay 5-day:Session 2	−0.054	0.141	−0.384	0.701
Delay 7-day:Session 2	0.02	0.133	0.151	0.88
Bias Question:Session 2	0.025	0.058	0.436	0.663
Delay 5-day:Bias Question:Session 2	0.001	0.141	0.007	0.995
Delay 7-day:Bias Question:Session 2	−0.029	0.134	−0.214	0.831

Open in a new tab

An anonymous reviewer asked if the ordering of the task (= the adaptation task first, the discrimination task second) would have affected the stability of behavioral responses over two sessions. It is possible that the adaptation happened in each day, which stabilized listeners’ discrimination performances. If we had tested their discrimination performances before the adaptation task, we could have seen a change of performances between sessions 1 and 2.

To address this, we conducted an additional data collection with 56 participants, using the statement-biasing, 3-day delay condition. We reversed the order of the two tasks while keeping everything else the same as the corresponding condition reported in the main manuscript. The results replicated the original ones (Figure A3), which suggests that the ordering of the two tasks were unlikely to have a large impact on the stability of the discrimination task performances we reported in the manuscript.

Figure A3. — A new group of participants (N=56) were recruited to complete the Statement-biasing condition with a 3-day delay in between. The task order was reversed, and they completed the discrimination task before the adaptation task on each day. The results overall replicated what we reported in 3.1 and 3.2 in the main body of the manuscript.

Appendix III. Question- and statement-biasing conditions comparison

Table A4.

Model summary for both bias conditions. Fixed effects estimates of the generalized linear mixed-effects are derived for bias, delay, session, block, continuum, as well as their interactions.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	0.627	0.039	15.942	<2e-16	***
Bias Question (vs. Statement)	−0.18	0.039	−4.593	4.37e-06	***
Delay 5-day (vs. 3-day)	−0.271	0.099	−2.727	0.006	**
Delay 7-day (vs. 5-day)	0.152	0.093	1.629	0.103
Session 2	0.209	0.039	5.319	1.04e-07	***
Block Post (vs. Pre)	0.277	0.022	12.627	<2e-16	***
Continuum	0	0.01	0.017	0.986
Bias Question:Delay 5-day	0.086	0.099	0.87	0.385
Bias Question:Delay 7-day	0.046	0.093	0.496	0.62
Bias Question:Session 2	−0.026	0.039	−0.665	0.506
Delay 5-day:Session 2	−0.042	0.099	−0.421	0.674
Delay 7-day:Session 2	0.013	0.093	0.138	0.891
Bias Question:Block Post	0.09	0.022	4.149	3.34e-05	***
Delay 5-day:Block Post	−0.032	0.055	−0.58	0.562
Delay 7-day:Block Post	−0.003	0.051	−0.057	0.955
Session 2:Block Post	−0.137	0.022	−6.299	2.88e-10	***
Bias Question:Continuum	1.281	0.01	122.914	<2e-16	***
Delay 5-day:Continuum	0.079	0.025	3.19	0.001	***
Delay 7-day:Continuum	0.011	0.022	0.493	0.622
Session 2:Continuum	−0.008	0.01	−0.826	0.409
Block Post:Continuum	−0.038	0.009	−4.051	5.10e-05	***
Bias Question:Delay 5-day:Session 2	−0.052	0.099	−0.528	0.598
Bias Question:Delay 7-day:Session 2	0.014	0.093	0.145	0.885
Bias Question:Delay 5-day:Block Post	−0.066	0.055	−1.192	0.233
Bias Question:Delay 7-day:Block Post	0.118	0.051	2.298	0.022	*
Bias Question:Session 2:Block Post	0.069	0.022	3.17	0.002	**
Delay 5-day:Session 2:Block Post	0.086	0.055	1.55	0.121
Delay 7-day:Session 2:Block Post	0.022	0.051	0.426	0.67
Bias Question:Delay 5-day:Continuum	−0.174	0.025	−7.056	1.71e-12	***
Bias Question:Delay 7-day:Continuum	0.087	0.022	3.958	7.56e-05	***
Bias Question:Session 2:Continuum	−0.016	0.01	−1.597	0.11
Delay 5-day:Session 2:Continuum	0.091	0.025	3.673	2.40e-4	***
Delay 7-day:Session 2:Continuum	−0.009	0.022	−0.427	0.669
Bias Question:Block Post:Continuum	−0.036	0.01	−3.465	0.001	***
Delay 5-day:Block Post:Continuum	0.09	0.024	3.792	1.49e-4	***
Delay 7-day:Block Post:Continuum	−0.039	0.021	−1.826	0.068
Session 2:Block Post:Continuum	0.023	0.009	2.444	0.015	*
Bias Question:Delay 5-day:Session 2:Block Post	−0.023	0.055	−0.412	0.68
Bias Question:Delay 7-day:Session 2:Block Post	−0.025	0.051	−0.491	0.623
Bias Question:Delay 5-day:Session 2:Continuum	−0.048	0.025	−1.953	0.051
Bias Question:Delay 7-day:Session 2:Continuum	−0.008	0.022	−0.365	0.715
Bias Question:Delay 5-day:Block Post:Continuum	−0.054	0.024	−2.264	0.024	*
Bias Question:Delay 7-day:Block Post:Continuum	0.045	0.021	2.151	0.032	*
Bias Question:Session 2:Block Post:Continuum	−0.002	0.009	−0.232	0.817
Delay 5-day:Session 2:Block Post:Continuum	−0.015	0.024	−0.608	0.543
Delay 7-day:Session 2:Block Post:Continuum	0.014	0.021	0.664	0.507
Bias Question:Delay 5-day:Session 2:Block Post:Continuum	0.023	0.024	0.954	0.34
Bias Question:Delay 7-day:Session 2:Block Post:Continuum	−0.011	0.021	−0.533	0.594

Open in a new tab

Appendix IV. Response accuracy and reaction times in the exposure phase

Figure A4. — Percent correct responses for each trial in order in the exposure phase by delay, bias, and session. Panel A shows data from the 3-day delay, B from the 5-day, and C from the 7-day. The left column is the question-biasing condition, and the right is the statement-biasing. The top row is Session 1; the bottom is Session 2. Red points and lines indicate trials that should be considered questions, and blue points and lines indicate trials that should be considered statements.

Appendix V. Mixed effect logistic regression for exposure phase responses

We constructed two separate generalized linear mixed-effect logistic regression models to predict the accuracy of responses (1 = correct, 0 = incorrect) in the question-biasing and in the statement-biasing conditions. In each, we tested effects of five fixed effect variables: (a) delay (factorial, sliding difference coded), (b) session (contrast-coded, session 1 = −1, session 2 = 1), (c) stimulus ambiguity (sum-coded, unambiguous = −1 vs. ambiguous = 1), (d) trial number (numerical, centered) and the interaction between them. The final models included the identical maximum random effect structure: by-subject intercepts and slopes for the interaction between delay, session, and stimulus.

Table A5.

Model summary for the exposure phase accuracy in the question-biasing condition. Fixed effects estimates of the generalized linear mixed-effects are derived for delay, session, stimulus, trial number, as well as their interactions.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	2.561	0.123	20.83	< 2e-16	***
Delay 5-day (vs 3-day)	−0.015	0.232	−0.064	0.949
Delay 7-day (vs. 5-day)	0.186	0.223	0.833	0.405
Session 2 (vs. 1)	0.017	0.11	0.151	0.88
Stimulus Ambiguous (vs. Unambiguous)	−1.63	0.101	−16.159	< 2e-16	***
Order	0.04	0.007	5.738	9.57e-09	***
Delay 5-day:Session 2	0.408	0.232	1.756	0.079
Delay 7-day:Session 2	−0.317	0.223	−1.423	0.155
Delay 5-day:Stimulus Ambiguous	−0.144	0.195	−0.736	0.462
Delay 7-day:Stimulus Ambiguous	−0.013	0.189	−0.069	0.945
Session 2:Stimulus Ambiguous	−0.01	0.088	−0.118	0.906
Delay 5-day:Order	−0.003	0.012	−0.267	0.79
Delay 7-day:Order	0.009	0.012	0.751	0.452
Session 2:Order	0	0.006	−0.068	0.946
Stimulus Ambiguous:Order	0.055	0.005	11.898	< 2e-16	***
Delay 5-day:Session 2:Stimulus Ambiguous	−0.154	0.195	−0.789	0.43
Delay 7-day:Session 2:Stimulus Ambiguous	0.08	0.189	0.426	0.67
Delay 5-day:Session 2:Order	−0.02	0.012	−1.656	0.098
Delay 7-day:Session 2:Order	0.012	0.012	1.047	0.295
Delay 5-day:Stimulus Ambiguous:Order	0.005	0.01	0.531	0.596
Delay 7-day:Stimulus Ambiguous:Order	−0.002	0.01	−0.245	0.806
Session 2:Stimulus Ambiguous:Order	0	0.004	0.08	0.936
Delay 5-day:Session 2:Stimulus Ambiguous:Order	0.005	0.01	0.513	0.608
Delay 7-day:Session 2:Stimulus Ambiguous:Order	−0.003	0.01	−0.294	0.769

Open in a new tab

Table A6.

Model summary for the exposure phase accuracy in the statement-biasing condition. Fixed effects estimates of the generalized linear mixed-effects are derived for delay, session, stimulus, trial number, as well as their interactions.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	2.652	0.122	21.774	< 2e-16	***
Delay 5-day (vs 3-day)	−0.061	0.234	−0.262	0.793
Delay 7-day (vs. 5-day)	−0.06	0.216	−0.28	0.78
Session 2 (vs. 1)	0.256	0.116	2.206	0.027
Stimulus Ambiguous (vs. Unambiguous)	−0.719	0.103	−6.974	3.07e-12	***
Order	0.055	0.008	7.072	1.53e-12	***
Delay 5-day:Session 2	−0.151	0.234	−0.644	0.52
Delay 7-day:Session 2	−0.21	0.216	−0.974	0.33
Delay 5-day:Stimulus Ambiguous	0.121	0.208	0.584	0.559
Delay 7-day:Stimulus Ambiguous	−0.116	0.19	−0.612	0.54
Session 2:Stimulus Ambiguous	0.528	0.098	5.364	8.16e-08	***
Delay 5-day:Order	−0.017	0.014	−1.217	0.223
Delay 7-day:Order	0.005	0.013	0.414	0.679
Session 2:Order	−0.013	0.007	−1.833	0.067
Stimulus Ambiguous:Order	0.042	0.006	7.274	3.48e-13	***
Delay 5-day:Session 2:Stimulus Ambiguous	−0.005	0.208	−0.023	0.981
Delay 7-day:Session 2:Stimulus Ambiguous	−0.101	0.19	−0.534	0.594
Delay 5-day:Session 2:Order	0.007	0.014	0.468	0.64
Delay 7-day:Session 2:Order	0.017	0.013	1.292	0.196
Delay 5-day:Stimulus Ambiguous:Order	−0.018	0.012	−1.466	0.143
Delay 7-day:Stimulus Ambiguous:Order	−0.002	0.01	−0.214	0.831
Session 2:Stimulus Ambiguous:Order	−0.022	0.005	−4.755	1.98e-06	***
Delay 5-day:Session 2:Stimulus Ambiguous:Order	0.004	0.012	0.318	0.751
Delay 7-day:Session 2:Stimulus Ambiguous:Order	0.007	0.01	0.684	0.494

Open in a new tab

We then constructed a combined model that includes data from both biasing conditions. In addition to the fixed effects described above, we added bias (factorial, contrast coded, statement-biasing = 1, question-biasing = −1) and its interaction with the other five fixed effect variables. The random effect structure included: by-subject intercepts and slopes for the interaction between delay, session, and stimulus.

Table A7.

Model summary for the exposure phase accuracy in the combined dataset. Fixed effects estimates of the generalized linear mixed-effects are derived for delay, session, stimulus, trial number, as well as their interactions. Rows highlighted in blue are significant at the α = 0.05 level.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	2.505	0.081	30.873	< 2e-16	***
Delay 5-day (vs. 3-day)	−0.037	0.157	−0.236	0.813
Delay 7-day (vs. 5-day)	0.079	0.148	0.534	0.593
Bias Statement (vs. Question)	0.061	0.063	0.969	0.333
Session 2 (vs. 1)	0.112	0.077	1.45	0.147
Stimulus Ambiguous (vs. Unambiguous)	−1.133	0.069	−16.386	< 2e-16	***
Order	0.051	0.005	9.779	< 2e-16	***
Delay 5-day:Bias Statement	0.019	0.157	0.119	0.906
Delay 7-day:Bias Statement	−0.174	0.148	−1.174	0.24
Delay 5-day:Session 2	0.116	0.157	0.738	0.461
Delay 7-day:Session 2	−0.219	0.148	−1.482	0.138
Bias Statement:Session 2	0.108	0.063	1.728	0.084
Delay 5-day:Stimulus Ambiguous	0.001	0.138	0.004	0.997
Delay 7-day:Stimulus Ambiguous	−0.099	0.13	−0.76	0.447
Bias Statement:Stimulus Ambiguous	0.373	0.055	6.77	1.29e-11	***
Session 2:Stimulus Ambiguous	0.248	0.065	3.83	0.000128	***
Delay 5-day:Order	−0.01	0.009	−1.064	0.287
Delay 7-day:Order	0.006	0.009	0.72	0.472
Bias Statement:Order	0.009	0.004	2.393	0.017	*
Session 2:Order	−0.006	0.005	−1.264	0.206
Stimulus Ambiguous:Order	0.048	0.004	13.325	< 2e-16	***
Delay 5-day:Bias Statement:Session 2	−0.253	0.157	−1.614	0.107
Delay 7-day:Bias Statement:Session 2	0.027	0.148	0.18	0.857
Delay 5-day:Bias Statement:Stimulus Ambiguous	0.119	0.138	0.862	0.389
Delay 7-day:Bias Statement:Stimulus Ambiguous	−0.008	0.13	−0.062	0.951
Delay 5-day:Session 2:Stimulus Ambiguous	−0.066	0.138	−0.476	0.634
Delay 7-day:Session 2:Stimulus Ambiguous	−0.042	0.13	−0.32	0.749
Bias Statement:Session 2:Stimulus Ambiguous	0.213	0.055	3.872	< 0.001	***
Delay 5-day:Bias Statement:Order	−0.008	0.009	−0.813	0.416
Delay 7-day:Bias Statement:Order	0	0.009	0.044	0.965
Delay 5-day:Session 2:Order	−0.006	0.009	−0.618	0.537
Delay 7-day:Session 2:Order	0.013	0.009	1.475	0.14
Bias Statement:Session 2:Order	−0.006	0.004	−1.504	0.133
Delay 5-day:Stimulus Ambiguous:Order	−0.007	0.008	−0.86	0.39
Delay 7-day:Stimulus Ambiguous:Order	−0.001	0.007	−0.083	0.934
Bias Statement:Stimulus Ambiguous:Order	−0.006	0.003	−2.077	0.038	*
Session 2:Stimulus Ambiguous:Order	−0.011	0.003	−3.543	< 0.001	***
Delay 5-day:Bias Statement:Session 2:Stimulus Ambiguous	0.056	0.138	0.407	0.684
Delay 7-day:Bias Statement:Session 2:Stimulus Ambiguous	−0.066	0.13	−0.508	0.612
Delay 5-day:Bias Statement:Session 2:Order	0.013	0.009	1.42	0.156
Delay 7-day:Bias Statement:Session 2:Order	0.003	0.009	0.326	0.744
Delay 5-day:Bias Statement:Stimulus Ambiguous:Order	−0.011	0.008	−1.447	0.148
Delay 7-day:Bias Statement:Stimulus Ambiguous:Order	−0.002	0.007	−0.287	0.774
Delay 5-day:Session 2:Stimulus Ambiguous:Order	0.005	0.008	0.586	0.558
Delay 7-day:Session 2:Stimulus Ambiguous:Order	0.003	0.007	0.41	0.682
Bias Statement:Session 2:Stimulus Ambiguous:Order	−0.012	0.003	−3.784	< 0.002	***
Delay 5-day:Bias Statement:Session 2:Stimulus Ambiguous:Order	0.001	0.008	0.102	0.919
Delay 7-day:Bias Statement:Session 2:Stimulus Ambiguous:Order	0.004	0.007	0.531	0.596

Open in a new tab

Appendix VI. Exploratory analyses on the inferred age/location and adaptation responses

To explore possible roles of inferred age and geographic regions, we conducted additional mixed-effects regression analyses. Note that these are post-hoc analyses to explore possible relationships between listeners’ social inferences and their adaptive understanding of intonation. However, the task was not designed to test this relationship and thus was not optimized to support the full model needed for a conclusive test. It is also worth noting that we focus on the statement-biasing condition, where the “ambiguous” contours may sound similar to what is often referred to as “uptalk”. Thus, the question was whether the inferred age and geographic location would predict the degree of adaptive shift that listeners showed after exposure. Specifically, we expected that (1) the younger age estimate and (2) the choice of “West” would be more likely to support the uptalk interpretation (= “statement” responses to ambiguous tokens). If so, these choices could predict more robust adaptive shift in each session in the statement-biasing condition.

1. Age

We used the log-transformed inferred age as a fixed effect to predict the adaptation before and after the exposure. As in the main model reported in section 3.3.2, the dependent variable is the trial level choice of the “statement” response. The largest model with delay × session × block × continuum × inferred age did not converge even with the simplest random effect structure. We therefore simplified the model by removing the delay conditions, based on the finding that the three delay conditions did not yield any significant differences. Thus, the model we report on here included four fixed effects (session × block × continuum × inferred age) and random intercepts for participants. Models with larger random effect structures failed to converge. The model summary is given in Table A8. As seen in the model reported in the Results section of the manuscript, the main effects of session, block, and continuum are all significant: Participants provided more statement responses toward the lower end of the continuum, did so more in the post-test than in the pre-test, and did so more in session 2 than in session 1. The key negative two-way interaction between block and session was highly significant, suggesting that adaptation was diminished in session 2 due to the maintained shift that already biased the pretest responses toward “statement.”

Table A8.

Model summary for the exploratory analysis using the inferred age of the talker as a predictor. Dependent variable is trial-level statement responses. Fixed effects estimates of the generalized linear mixed-effects are derived for session, block, continuum, inferred age as well as their interactions. Rows highlighted in blue are significant at the α = 0.05 level.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	0.762	0.048	15.938	0	***
Session 2 (vs. 1)	0.219	0.048	4.59	0	***
Block Post (vs. Pre)	0.186	0.021	8.988	0	***
Continuum	−1.214	0.013	−90.756	0	***
LogAgeEstimate	−0.338	0.289	−1.17	0.242
Session 2:Block Post	−0.186	0.021	−8.98	0	***
Session 2:Continuum	0.021	0.013	1.623	0.105
Block Post:Continuum	−0.013	0.012	−1.121	0.262
Session 2:LogAgeEstimate	−0.242	0.289	−0.837	0.402
Block Post:LogAgeEstimate	−0.062	0.12	−0.517	0.605
Continuum:LogAgeEstimate	0.526	0.072	7.356	0	***
Session 2:Block Post:Continuum	0.026	0.012	2.198	0.028	*
Session 2:Block Post:LogAgeEstimate	0.072	0.12	0.6	0.548
Session 2:Continuum:LogAgeEstimate	−0.045	0.072	−0.63	0.529
Block Post:Continuum:LogAgeEstimate	−0.186	0.065	−2.854	0.004	**
Session 2:Block Post:Continuum:LogAgeEstimate	0.027	0.065	0.417	0.677

Open in a new tab

Critically, neither the main effect of inferred age nor its interaction with session × block was significant. This means that the impression of the talker’s age did not change the overall likelihood of the “statement” response or its adaptive shift after exposure. In addition, the continuum × age interaction was highly significant. This means that higher age (= unlikely uptalk user) predicted more statement responses for the higher end of the continuum, which is the opposite of what we expected.

Interestingly, however, the negative three-way interaction including block, continuum, and inferred age was significant. This could mean that the amount of adaptation in each session was smaller when the inferred age was higher (= unlikely uptalk user), especially at the upper end of the continuum where the pre-post difference would be expected to be more pronounced. (The lower end of the continuum is capped by the floor effect.) This may support the idea that the expectation of the talker being an uptalk user can change how readily listeners shift their categorization judgments. However, given the exploratory nature of the analysis, this is not a definitive conclusion and requires further investigation.

2. Geographical region

We recoded the answers to the geographical location as West vs. Other. This was because we did not have a clear hypothesis about how listeners’ choice of other geographic regions would affect their question vs. statement judgments. Here, as above, the dependent variable is the trial-level choice of the “statement” response. And the model included four fixed effects (session × block × continuum × inferred region = West) and random intercepts and slopes by participant for session, block, and continuum. Models with larger random effects structures did not converge.

The model summary is given in Table A9. As seen in the model reported in the Results section of the manuscript, the main effects of session, block, and continuum are all significant: participants provided more statement responses toward the lower end of the continuum, did so more in the post-test than in the pre-test, and did so more in session 2 than in session 1. The key negative two-way interaction between block and session was highly significant, suggesting that adaptation was diminished in session 2 due to the maintained shift that already biased the pretest responses toward “statement.”

Table A9.

Model summary for the exploratory analysis using the inferred region of the talker as a predictor. Fixed effects estimates of the generalized linear mixed-effects are derived for session, block, continuum, inferred region as well as their interactions. Rows highlighted in blue are significant at the α = 0.05 level.

	Estimate	Std. Error	z value	Pr(>\|z\|)
(Intercept)	1.009	0.062	16.343	0	***
Session 2 (vs. 1)	0.361	0.061	5.883	0	***
Block Post (vs. Pre)	0.209	0.039	5.382	0	***
Continuum	−1.581	0.034	−46.94	0	***
Region West (vs. other)	0.191	0.06	3.153	0.002	**
Session 2:Block Post	−0.229	0.038	−6.028	0	***
Session 2:Continuum	−0.019	0.031	−0.629	0.53
Block Post:Continuum	0.014	0.015	0.944	0.345
Session 2:Region West	−0.009	0.06	−0.144	0.885
Block Post:Region West	−0.056	0.038	−1.468	0.142
Continuum:Region West	−0.093	0.03	−3.058	0.002	**
Session 2:Block Post:Continuum	0.031	0.015	2.114	0.035	*
Session 2:Block Post:Region West	−0.044	0.038	−1.163	0.245
Session 2:Continuum:Region West	−0.055	0.03	−1.824	0.068
Block Post:Continuum:Region West	0.009	0.014	0.651	0.515
Session 2:Block Post:Continuum:Region West	0.004	0.014	0.273	0.785

Open in a new tab

The main effect of inferred region was also significant. This suggests that those who thought the talker was from the West were overall more likely to respond “statement” than those who chose other regions. This is in line with the expectation that participants who inferred the talker to be a likely uptalk user would consider a wider range of contours as “statements”. In addition, the continuum × region interaction was significant with a negative coefficient. This indicates that those who thought the talker was from the West (=likely uptalk user) were not responding “statement” as much when the contour had a clear rise. This is also consistent with the expectation that the effect of inferred uptalk status would be most pronounced for ambiguous tokens in the stimuli.

However, as in the model with inferred age above, no other interactions including block × inferred region were significant. This means that the inferred regional origin of the talker did not affect the amount of adaptation in each session or maintaining across sessions.

Appendix VII. Comparisons of exposure and test stimulus features

Our test and exposure stimuli were created based on separate sets of base recordings. Consequently, the six 11-step continua (five exposure and one test) we used were not identical in their critical phonetic-prosodic features. In particular, the exposure items included a long vowel, and a part of the speaker’s encoding strategies included changing the final syllable duration while keeping pitch differences less drastic as compared to the test items. Figure A5 below overlays the pitch contours of the exposure stimuli on top of the test stimuli. The exposure stimuli exhibited variability, and none matched straightforwardly to the test stimuli. This suggests that listeners’ “adaptation” within a single session included some form of generalization or transposition from the exposure to the test stimuli. The maintenance therefore must have somehow operated over the outcome of this generalization/transposition.

Figure A5. — Pitch contours of the exposure (colored) and test (gray) stimuli.

References

Anderson MC, & Hulbert JC (2021). Active forgetting: Adaptation of memory by prefrontal control. Annual Review of Psychology, 72, 1–36. 10.1146/annurev-psych-072720-094140 [DOI] [PubMed] [Google Scholar]
Apfelbaum KS, & McMurray B (2015). Relative cue encoding in the context of sophisticated models of categorization: Separating information from categorization. Psychonomic Bulletin & Review, 22(3), 916–943. [DOI] [PMC free article] [PubMed] [Google Scholar]
Apfelbaum KS, Kutlu E, McMurray B, Kapnoula EC. (2022). Don’t force it! Gradient speech categorization calls for continuous categorization tasks. J Acoust Soc Am. 152 (6):3728. doi: 10.1121/10.0015201. [DOI] [PMC free article] [PubMed] [Google Scholar]
Arnhold A, & Kyröläinen A-J (2017). Modelling the interplay of multiple cues in prosodic focus marking. Laboratory Phonology, 8(1). 10.5334/labphon.78 [DOI] [Google Scholar]
Arnold JE, & Watson DG (2015). Synthesizing meaning and processing approaches to prosody: Performance matters. Language, Cognition and Neuroscience, 30(1-2), 88–102. 10.1080/01690965.2013.840733 [DOI] [PMC free article] [PubMed] [Google Scholar]
Arthur SE, Wise RJS, Thomas DL, & Leech R (2021). The neural dynamics underlying continuous speech comprehension in naturalistic conditions. Communications Biology, 4(1), 1094. 10.1038/s42003-021-02618-z34535759 [DOI] [Google Scholar]
Arvaniti A (2019). Crosslinguistic variation, phonetic variability, and the formation of categories in intonation. In Calhoun S, Escudero P, Tabain M, & Warren P (Eds.), Proceedings of the 19th International Congress of Phonetic Sciences, Melbourne, Australia 2019 (pp. 1–6). [Google Scholar]
Arvaniti A, & Garding G (2007). Dialectal variation in the rising accents of American English. In Cole J & Hualde JI (Eds.), Laboratory Phonology 9 (pp. 547–576). Mouton de Gruyter. [Google Scholar]
Arvaniti A, Tilsen S, Cangemi F, & Shue Y-L (2024). Continuous and categorical aspects of intonational categories. Journal of Phonetics, 102, 101287. 10.1016/j.wocn.2023.101287 [DOI] [Google Scholar]
Aylett M, & Turk A (2004). The smooth signal redundancy hypothesis: A functional explanation for relationships between redundancy, prosodic prominence, and duration in spontaneous speech. Language and Speech, 47(1), 31–56. 10.1177/00238309040470010201 [DOI] [PubMed] [Google Scholar]
Bartle A, & Dellwo V (2015). Rhythmic variability and a repair effect on phonological development. In Proceedings of the 18th International Congress of Phonetic Sciences. University of Glasgow. [Google Scholar]
Baumann S, & Grice M (2006). The intonation of accessibility. Journal of Pragmatics, 38(10), 1636–1657. 10.1016/j.pragma.2005.03.017 [DOI] [Google Scholar]
Bent T, & Baese-Berk MM (2021). Listening to non-native speech: Examining the roles of linguistic and indexical information. The Journal of the Acoustical Society of America, 149(2), 1369–1379. [Google Scholar]
Boersma P, & Weenink D (2012). Praat: Doing phonetics by computer [Computer program]. Version 5.3.23. http://www.praat.org/
Bolinger D (1986). Intonation and its parts: Melody in spoken English. Stanford University Press. [Google Scholar]
Bosker HR (2021). Accounting for L1 and L2 listening: Investigating the time-course of speaker-specific processing with reaction times. Language Learning, 71(S1), 38–70. [Google Scholar]
Bosker HR (2022). Prosodic processing in a second language does not differ between information structure-based and acoustic-phonetic based learners. Bilingualism: Language and Cognition, 25(4), 689–700. 10.1017/S1366728922000281 [DOI] [Google Scholar]
Bradlow AR, Bassard AM, & Paller KA (2023). Generalized perceptual adaptation to second-language speech: Variability, similarity, and intelligibility. The Journal of the Acoustical Society of America, 154(3), 1601–1613. 10.1121/10.0020914 [DOI] [PMC free article] [PubMed] [Google Scholar]
Brown-Schmidt S, Yoon SO, & Ryskin RA (2015). People as contexts in conversation. Psychology of Learning and Motivation, 62, 59–99. 10.1016/bs.plm.2014.09.003 [DOI] [Google Scholar]
Buxó-Lugo A, Toscano JC, & Watson DG (2018). Effects of participant engagement on prosodic prominence. Discourse Processes, 55(3), 305–323. [DOI] [PMC free article] [PubMed] [Google Scholar]
Calhoun S, & White J (2025). Modeling prosodic prominence in second language acquisition. Language Learning, 75(1), 1–34. 10.1111/lang.12619 [DOI] [Google Scholar]
Cangemi F, & Grice M (2016). The importance of a distributional approach to categoriality in autosegmental-metrical accounts of intonation. Laboratory Phonology, 7(1). 10.5334/labphon.28 [DOI] [Google Scholar]
Cangemi F, Krüger M, & Grice M (2015). Listener-specific perception of speaker-specific productions in intonation. In Individual Differences in Speech Production and Perception (pp. 123–145). Peter Lang. [Google Scholar]
Carvalho PF, Chen CH, & Yu C (2021). The distributional properties of exemplars affect category learning and generalization. Scientific Reports, 11(1), 1–10. 10.1038/s41598-021-90743-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
Chodroff E, & Cole JS (2019). Testing the distinctiveness of intonational tunes: Evidence from imitative productions in American English. In Interspeech 2019 (pp. 3109–3113). 10.21437/interspeech.2019-2684 [DOI] [Google Scholar]
Chodroff E, & Wilson C (2020). Perceptual normalization of voice pitch depends on between-talker variability. Attention, Perception, & Psychophysics, 82(7), 3294–3313. [Google Scholar]
Clarke-Davidson CM, Luce PA, & Sawusch JR (2008). Does perceptual learning in speech reflect changes in phonetic category representation or decision bias? Perception & Psychophysics, 70(4), 604–618. [DOI] [PubMed] [Google Scholar]
Clayards M, Tanenhaus MK, Aslin RN, & Jacobs RA (2008). Perception of speech reflects optimal use of probabilistic speech cues. Cognition, 108(3), 804–809. 10.1016/j.cognition.2008.04.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
Clopper CG, & Smiljanic R (2011). Effects of gender and regional dialect on prosodic patterns in American English. Journal of Phonetics, 39(2), 237–245. 10.1016/j.wocn.2011.02.006 [DOI] [PMC free article] [PubMed] [Google Scholar]
Cole J (2015). Prosody in context: A review. Language, Cognition and Neuroscience, 30(1-2), 1–31. 10.1080/23273798.2014.963130 [DOI] [Google Scholar]
Cummings SN, Hodges E, & Theodore RM (in press). Cumulative input sensitivity predicts both attenuation and stability of lexically guided perceptual learning. Psychonomic Bulletin & Review. [DOI] [PubMed] [Google Scholar]
Cummings SN (2025). Linking lexically guided perceptual learning to statistical patterns in speech input [Doctoral dissertation, University of Connecticut]. [Google Scholar]
Cummings SN, & Theodore RM (2022). Perceptual learning of multiple talkers: Determinants, characteristics, and limitations. Attention, Perception, & Psychophysics, 84(7), 2335–2359. 10.3758/s13414-022-02556-6 [DOI] [PubMed] [Google Scholar]
Cutler A, & Swinney DA (1987). Prosody and the development of comprehension. Journal of Child Language, 14(1), 145–167. 10.1017/S0305000900012782 [DOI] [PubMed] [Google Scholar]
Cychosz M, & Newman RS (2023). Perceptual normalization for speaking rate occurs below the level of the syllable. The Journal of the Acoustical Society of America, 153(3), 1486. 10.1121/10.0017360 [DOI] [PMC free article] [PubMed] [Google Scholar]
DeKeyser R (2020). Skill acquisition theory. In VanPatten B, Keating GD, & Wulff S (Eds.), Theories in Second Language Acquisition: An Introduction (3rd ed., pp. 83–104). Routledge. [Google Scholar]
Dellwo V, Leemann A, & Kolly M-J (2015). Rhythmic variability between speakers: Articulatory, prosodic, and linguistic factors. The Journal of the Acoustical Society of America, 137(3), 1513–1528. 10.1121/1.4906837 [DOI] [PubMed] [Google Scholar]
Drouin JR, & Theodore RM (2019). Lexically guided perceptual learning is robust to task-based changes in listening strategy. The Journal of the Acoustical Society of America, 144(2), 1089. [DOI] [PMC free article] [PubMed] [Google Scholar]
Earle FS, & Qi Z (2022). Understanding the role of sleep in memory consolidation: A focus on speech perception. Brain and Language, 228, 105095. 10.1016/j.bandl.2022.105095 [DOI] [PubMed] [Google Scholar]
Earle FS, & Ullman MT (2021). Deficits of learning in procedural memory and consolidation in declarative memory in adults with developmental language disorder. Journal of Speech, Language, and Hearing Research, 64(2), 531–541. 10.1044/2020_JSLHR-20-00292 [DOI] [PMC free article] [PubMed] [Google Scholar]
Eisner F, & McQueen JM (2006). Perceptual learning in speech: Stability over time. The Journal of the Acoustical Society of America, 119(4), 1950–1953. 10.1121/1.2178721 [DOI] [PubMed] [Google Scholar]
Fenn KM, Nusbaum HC, & Margoliash D (2003). Consolidation during sleep of perceptual learning of spoken language. Nature, 425(6958), 614–616. 10.1038/nature01951 [DOI] [PubMed] [Google Scholar]
Gnanateja GN, Grandhi SA, & Cole J (2025). Context-dependent prosodic variation in speech production and perception. Laboratory Phonology, 16(1). 10.16995/labphon.10067 [DOI] [Google Scholar]
Goldinger SD (1998). Echoes of echoes? An episodic theory of lexical access. Psychological Review, 105(2), 251–279. [DOI] [PubMed] [Google Scholar]
Goldrick M, & Cole J (2023). Advancement of phonetics in the 21st century: Exemplar models of speech production. Journal of Phonetics, 99, 101254. 10.1016/j.wocn.2023.101254 [DOI] [Google Scholar]
Goudbeek M, Swingley D, & Smits R (2009). Supervised and unsupervised learning of multidimensional acoustic categories. Journal of Experimental Psychology: Human Perception and Performance, 35(6), 1913–1933. 10.1037/a0015781 [DOI] [PMC free article] [PubMed] [Google Scholar]
Grice M, Ritter S, Niemann H, & Roettger TB (2017). Integrating the discreteness and continuity of intonational categories. Journal of Phonetics, 64, 90–107. 10.1016/j.wocn.2017.03.003 [DOI] [Google Scholar]
Gryllia S, Baltazani M, & Arvaniti A (2018). The role of pragmatics and politeness in explaining prosodic variability. In Proceedings of the 9th International Conference on Speech Prosody 2018 (pp. 32–36). [Google Scholar]
Gunlogson C (2008). A question of commitment. Belgian Journal of Linguistics, 22(1), 101–136. 10.1075/bjl.22.06gun [DOI] [Google Scholar]
Hansen M, Horne M, Lindström E, & Roll M (2015). Neurocognitive processes of prosodic sentence type recognition in L2. In Babatsouli E & Ingram D (Eds.), Proceedings of the International Symposium on Monolingual and Bilingual Speech 2015 (pp. 174–181). [Google Scholar]
Heffner CC, & Myers EB (2021). Individual differences in phonetic plasticity across native and nonnative contexts. Journal of Speech, Language, and Hearing Research, 64(9), 3720–3733. 10.1044/2021_JSLHR-20-00650 [DOI] [PubMed] [Google Scholar]
Heffner CC, Fuhrmeister P, Luthra S, Mechtenberg H, Saltzman D, & Myers EB (2022). Reliability and validity for perceptual flexibility in speech. Brain and Language, 226, 105070. 10.1016/j.bandl.2021.105070 [DOI] [PMC free article] [PubMed] [Google Scholar]
Hitczenko K, & Feldman NH (2022). Naturalistic speech supports distributional learning across contexts. Proceedings of the National Academy of Sciences, 119(38), e2123230119. 10.1073/pnas.2123230119 [DOI] [PMC free article] [PubMed] [Google Scholar]
Holliday JJ (2021). Context effects on vowel-to-vowel coarticulation in three dialects of American English. Language and Speech, 64(3), 694–720. 10.1177/0023830920951895 [DOI] [Google Scholar]
Huttenlauch C, De Beer C, Hanne S, & Wartenburger I (2021). Production of prosodic cues in coordinate name sequences addressing varying interlocutors. Laboratory Phonology: Journal of the Association for Laboratory Phonology, 12(1), 1. 10.5334/labphon.221 [DOI] [Google Scholar]
Idemaru K, & Holt LL (2011). Word recognition reflects dimension-based statistical learning. Journal of Experimental Psychology: Human Perception and Performance, 37(6), 1939–1956. [DOI] [PMC free article] [PubMed] [Google Scholar]
Idemaru K, & Holt LL (2020). Talker-contingent auditory perceptual learning generalizes across voices and depends on early sensory processing. Attention, Perception, & Psychophysics, 82(1), 191–210. [Google Scholar]
Jesse A, & Kaplan E (2019). Reward sensitivity and perceptual learning in speech. Attention, Perception, & Psychophysics, 81(5), 1367–1384. 10.3758/s13414-018-01651-x [DOI] [PubMed] [Google Scholar]
Jessen M (2008). Forensic phonetics. Language and Linguistics Compass, 2(4), 671–711. 10.1111/j.1749-818X.2008.00066.x [DOI] [Google Scholar]
Johnson K (1997). Speech perception without speaker normalization: An exemplar model. In Johnson K & Mullennix JW (Eds.), Talker Variability in Speech Processing (pp. 145–165). Academic Press. [Google Scholar]
Johnson K, & Sjerps MJ (2021). Speaker normalization in speech perception. In Pardo JS, Nygaard LC, Remez RE, & Pisoni DB (Eds.), The Handbook of Speech Perception (2nd ed., pp. 145–176). Wiley. 10.1002/9781119184096.ch6 [DOI] [Google Scholar]
Jun S-A, & Bishop J (2015). Priming implicit prosody: Prosodic boundaries and individual differences. Language and Speech, 58(4), 459–473. 10.1177/0023830914563368 [DOI] [PubMed] [Google Scholar]
Kaschak MP, Kutta TJ, & Jones JL (2011). Structural priming as implicit learning: Cumulative priming effects and individual differences. Psychonomic Bulletin & Review, 18(6), 1133–1139. [DOI] [PMC free article] [PubMed] [Google Scholar]
Kleinschmidt DF, & Jaeger TF (2015). Robust speech perception: Recognize the familiar, generalize to the similar, and adapt to the novel. Psychological Review, 122(2), 148–203. 10.1037/a0038695 [DOI] [PMC free article] [PubMed] [Google Scholar]
Kronrod Y, Coppess E, & Feldman NH (2016). A unified model of categorical effects in consonant and vowel perception. Psychonomic Bulletin & Review, 23(6), 1681–1712. 10.3758/s13423-016-1049-y [DOI] [PubMed] [Google Scholar]
Kuhl PK, Conboy BT, Padden D, Nelson T, & Pruitt J (2010). Early speech perception and later language development: Implications for the critical period. Language Learning and Development, 1(4), 237–264. [Google Scholar]
Kurumada C, & Buxó-Lugo A (2024). Intonation adaptation to multiple talkers. Journal of Experimental Psychology: Learning, Memory, and Cognition, 50(12), 1954–1981. 10.1037/xlm0001419 [DOI] [PubMed] [Google Scholar]
Kurumada C, & Roettger TB (2022). Thinking probabilistically in the study of intonational speech prosody. WIREs Cognitive Science, 13(1), e1579. 10.1002/wcs.1579 [DOI] [PubMed] [Google Scholar]
Kurumada C, Xie X, & Buxó-Lugo A (2024). Talker-specific adaptation in prosodic processing: Investigating rapid perceptual adjustment to interlocutor variation. Cognition, 242, 105638.37839251 [Google Scholar]
Ladd DR (2008). Intonational Phonology (2nd ed.). Cambridge University Press. [Google Scholar]
Ladd DR, & Morton R (1997). The perception of intonational emphasis: Continuous or categorical? Journal of Phonetics, 25(3), 313–342. 10.1006/S0095-4470(97)90063-X [DOI] [Google Scholar]
Lavan N, Burton AM, Scott SK, & McGettigan C (2024). Voice identity processing in unfamiliar, newly learned, and familiar voices: A systematic review. Psychonomic Bulletin & Review, 31(1), 67–92. 10.3758/s13423-023-02333-9 [DOI] [Google Scholar]
Lee C-Y, Yen H-L, Yeh P-W, Lin W-H, Cheng Y-Y, Tzeng Y-L, & Wu H-C (2010). Fluent speech language representation in the brain: Effects of frequency and word class. Neuropsychologia, 48(10), 2988–3000. 10.1016/j.neuropsychologia.2010.06.008 [DOI] [Google Scholar]
Li J, Luo J, & Yu L (2021). Talker normalization of vowel categories in auditory cortex: Evidence from fMRI-based multivariate pattern analyses. The Journal of the Acoustical Society of America, 149(4), 2643–2654. [Google Scholar]
Liu L, & Holt LL (2015). Dimension-based statistical learning of vowels. Journal of Experimental Psychology: Human Perception and Performance, 41(6), 1783–1798. [DOI] [PMC free article] [PubMed] [Google Scholar]
Liu L, & Jaeger TF (2018). Inferring causes during speech perception. Cognition, 174, 55–70. 10.1016/j.cognition.2018.01.003 [DOI] [PMC free article] [PubMed] [Google Scholar]
Liu L, & Jaeger TF (2019). Talker-specific effects on spoken word recognition in immediate and delayed memory tests. Attention, Perception, & Psychophysics, 81(7), 2215–2233. [Google Scholar]
Lobanov BM (1971). Classification of Russian vowels spoken by different speakers. The Journal of the Acoustical Society of America, 49(2B), 606–608. [Google Scholar]
Lorenzen J, & Baumann S (2024). Adaptive prosody processing: Effects of prosody type, processing load, and adaptation direction. Journal of Phonetics, 103, 101311. 10.1016/j.wocn.2024.101311 [DOI] [Google Scholar]
Luo Y, Mitterer H, Zhou X, & Chen Y (2024). Flexibility and stability in lexical tone recalibration: Evidence from tone perceptual learning. Language and Speech, 00238309241291536. 10.1177/00238309241291536 [DOI] [PubMed] [Google Scholar]
Luthra S, Guediche S, Blumstein SE, & Myers EB (2021). Neural substrates of subphonemic variation and lexical competition in spoken word recognition. Language, Cognition and Neuroscience, 36(2), 151–169. [DOI] [PMC free article] [PubMed] [Google Scholar]
Maddox WT, Chandrasekaran B, Smayda K, & Yi H-G (2013). Dual systems of speech category learning across the lifespan. Psychology and Aging, 28(4), 1042–1056. 10.1037/a0034969 [DOI] [PMC free article] [PubMed] [Google Scholar]
Malisz Z, Bröker F, & Andreeva B (2018). Dimensions of rhythmic variability in spontaneous speech. In Proceedings of the 9th International Conference on Speech Prosody 2018 (pp. 1054–1058). 10.21437/SpeechProsody.2018-213 [DOI] [Google Scholar]
Maslowski M, Meyer AS, & Bosker HR (2019). How the tracking of habitual rate influences speech perception. Journal of Experimental Psychology: Learning, Memory, and Cognition, 45(1), 128–138. [DOI] [PubMed] [Google Scholar]
McAuliffe MJ, & Babel M (2016). Stimulus-directed attention attenuates lexically-guided perceptual learning. The Journal of the Acoustical Society of America, 140(3), 1727–1738. [DOI] [PubMed] [Google Scholar]
McMurray B, & Jongman A (2011). What information is necessary for speech categorization? Harnessing variability in the speech signal by integrating cues computed relative to expectations. Psychological Review, 118(2), 219–246. 10.1037/a0022325 [DOI] [PMC free article] [PubMed] [Google Scholar]
McQueen JM, & Dilley LC (2020). Prosody and spoken-word recognition. In Gussenhoven C & Chen A (Eds.), The Oxford Handbook of Language Prosody (pp. 554–565). Oxford University Press. 10.1093/oxfordhb/9780198832232.013.33 [DOI] [Google Scholar]
Mechtenberg H, Cheng L, Luthra S, & Myers EB (2025). Cognitive load does not affect lexically-guided perceptual learning. Attention, Perception, & Psychophysics. [Google Scholar]
Milne AE, Petkov CI, & Wilson B (2021). Auditory and visual sequence learning in humans and monkeys using an artificial grammar learning paradigm. Neuroscience, 389, 205–217. 10.1016/j.neuroscience.2017.06.059 [DOI] [PMC free article] [PubMed] [Google Scholar]
Myers EB, & Mesite LM (2014). Neural systems underlying perceptual adjustment to non-standard speech tokens. Journal of Memory and Language, 76, 80–93. [DOI] [PMC free article] [PubMed] [Google Scholar]
Nakamura C, White L, & Kondo M (2021). Rate normalization in the perception of L2 prosody. The Journal of the Acoustical Society of America, 150(4), A82–A82. [Google Scholar]
Nearey TM (1989). Static, dynamic, and relational properties in vowel perception. The Journal of the Acoustical Society of America, 85(5), 2088–2113. [DOI] [PubMed] [Google Scholar]
Newman RS, & Sawusch JR (2009). Perceptual normalization for speaking rate III: Effects of the rate of one voice on perception of another. Journal of Phonetics, 37(1), 46–65. [DOI] [PMC free article] [PubMed] [Google Scholar]
Norris D, & McQueen JM (2008). Shortlist B: A Bayesian model of continuous speech recognition. Psychological Review, 115(2), 357–395. 10.1037/0033-295X.115.2.357 [DOI] [PubMed] [Google Scholar]
Norris D, McQueen JM, & Cutler A (2003). Perceptual learning in speech. Cognitive Psychology, 47(2), 204–238. [DOI] [PubMed] [Google Scholar]
Nota N, Trujillo J, & De Ruiter JP (2023). Variability in gesture production and perception. Language and Cognition, 15(2), 241–268. 10.1017/langcog.2022.31 [DOI] [Google Scholar]
Nusbaum HC, & Magnuson JS (1997). Talker normalization: Phonetic constancy as a cognitive process. In Johnson K & Mullennix JW (Eds.), Talker Variability in Speech Processing (pp. 109–132). Academic Press. [Google Scholar]
Nygaard LC, & Pisoni DB (1998). Talker-specific learning in speech perception. Perception & Psychophysics, 60(3), 355–376. 10.3758/BF03206860 [DOI] [PubMed] [Google Scholar]
Ostrow A, & Kurumada C (2025). Short-term and long-term adaptive changes in prosodic comprehension. In Proceedings of the 38th Annual Conference on Human Sentence Processing. University of Michigan. [Google Scholar]
Ostrow A & Kurumada C, (2025, April 18). Short-term and long-term adaptive changes in prosodic comprehension. OSF. https://osf.io/3xvsh/ [DOI] [PMC free article] [PubMed]
Patel R, Connaghan K, Franco D, Edsall E, Forgit D, Olsen L, & Russell S (2011). The Caterpillar: A novel reading passage for assessment of motor speech disorders. American Journal of Speech-Language Pathology, 22(1), 1–9. [DOI] [PubMed] [Google Scholar]
Persson F, & Jaeger TF (2023). Continuous adaptation in native and non-native speech perception: Implications for theories of adaptation. The Journal of the Acoustical Society of America, 153(6), 3368–3380. [Google Scholar]
Pickering MJ, & Branigan HP (1998). The representation of verbs: Evidence from syntactic priming in language production. Journal of Memory and Language, 39(4), 633–651. 10.1006/jmla.1998.2592 [DOI] [Google Scholar]
Pierrehumbert JB (2001). Exemplar dynamics: Word frequency, lenition and contrast. In Bybee J & Hopper P (Eds.), Frequency and the Emergence of Linguistic Structure (pp. 137–157). John Benjamins. [Google Scholar]
Pierrehumbert JB, & Hirschberg J (1990). The meaning of intonational contours in the interpretation of discourse. In Cohen PR, Morgan J, & Pollack ME (Eds.), Intentions in Communication (pp. 271–311). MIT Press. [Google Scholar]
Pisanski K, Cartei V, McGettigan C, Raine J, & Reby D (2017). Voice modulation: A window into the origins of human vocal control? Trends in Cognitive Sciences, 20(4), 304–318. 10.1016/j.tics.2016.01.002 [DOI] [PubMed] [Google Scholar]
Podesva RJ, & Callier P (2015). Voice quality and identity. Annual Review of Applied Linguistics, 35, 173–194. 10.1017/S0267190514000270 [DOI] [Google Scholar]
Quam C, & Creel SC (2021). Impacts of acoustic-phonetic variability on perceptual development for spoken language: A review. Wiley Interdisciplinary Reviews: Cognitive Science, 12(3), e1558. 10.1002/wcs.1558 [DOI] [PMC free article] [PubMed] [Google Scholar]
Quam C, Wang A, Maddox WT, Golisch K, & Lotto A (2018). Procedural-memory, working-memory, and declarative-memory skills are each associated with dimensional integration in sound-category learning. Frontiers in Psychology, 9, 1828. 10.3389/fpsyg.2018.01828 [DOI] [PMC free article] [PubMed] [Google Scholar]
R Core Team. (2024). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://www.R-project.org/s [Google Scholar]
Roettger TB, & Franke M (2019). Evidential strength of intonational cues and rational adaptation to (un-)reliable intonation. Cognitive Science, 43(7), e12745. 10.1111/cogs.12745 [DOI] [PubMed] [Google Scholar]
Roettger TB, & Rimland K (2020). Listeners rapidly adapt to new talker prosody: Evidence from adaptation priming. In Proceedings of the 10th International Conference on Speech Prosody 2020 (pp. 844–848). 10.21437/SpeechProsody.2020-172 [DOI] [Google Scholar]
Saltzman D, & Myers EB (2021). The time course of individual learning history effects on perceptual adjustment. Attention, Perception, & Psychophysics, 83(3), 989–1009. 10.3758/s13414-020-02150-x [DOI] [Google Scholar]
Samuel AG, & Dumay N (2021). Auditory selective adaptation moment by moment, at multiple timescales. Journal of Experimental Psychology: Human Perception and Performance, 47(4), 596–615. 10.1037/xhp0000841 [DOI] [PubMed] [Google Scholar]
Samuel AG, & Kraljic T (2009). Perceptual learning for speech. Attention, Perception, & Psychophysics, 71(6), 1207–1218. 10.3758/APP.71.6.1207 [DOI] [PubMed] [Google Scholar]
Severijnen GGA, Bosker HR, Plag I, & McQueen JM (2021). Listeners track talker-specific prosody to deal with talker-variability. Brain Research, 1755, 147477. 10.1016/j.brainres.2020.147477 [DOI] [PubMed] [Google Scholar]
Severijnen GGA, Di Dona G, Bosker HR, & McQueen JM (2023). Tracking talker-specific cues to lexical stress: Evidence from perceptual learning. Journal of Experimental Psychology: Human Perception and Performance, 49(4), 549–565. 10.1037/xhp0001105 [DOI] [PubMed] [Google Scholar]
Severijnen GGA, Bosker HR, & McQueen JM (2024a). Learning to use talker-specific cues to lexical stress: More rapidly when staying with the same talker. Attention, Perception, & Psychophysics, 86(1), 63–81. 10.3758/s13414-023-02789-3 [DOI] [Google Scholar]
Severijnen GGA, Bosker HR, & McQueen JM (2024b). Your “VOORnaam” is not my “VOORnaam”: An acoustic analysis of individual talker differences in word stress in Dutch. Journal of Phonetics 103, 101296. doi: 10.1016/j.wocn.2024.101296 [DOI] [Google Scholar]
Severijnen GGA, McQueen JM, & Bosker HR (2025). Cross-talker generalization of talker-specific phonetic variation in speech perception. Journal of Phonetics, 98, 101243. [Google Scholar]
Sharf DJ, & Ohde RN (1981). Physiologic, acoustic, and perceptual aspects of coarticulation: Implications for the remediation of articulatory disorders. In Lass NJ (Ed.), Speech and Language: Advances in Basic Research and Practice (Vol. 5, pp. 153–247). Academic Press. [Google Scholar]
Sjerps MJ, Fox NP, Johnson K, & Chang EF (2019). Speaker-normalized sound representations in the human auditory cortex. Nature Communications, 10(1), 2465. 10.1038/s41467-019-10365-z [DOI] [PMC free article] [PubMed] [Google Scholar]
Steffman J, & Cole J (2024). Acoustic and perceptual evidence for systematic variation in English intonation. Phonology, 41(1), 1–35. 10.1017/S0952675723000271 [DOI] [Google Scholar]
Steffman J, Cole J, & Shattuck-Hufnagel S (2024). Intonational categories and continua in American English rising nuclear tunes. Journal of Phonetics, 104, 101310. 10.1016/j.wocn.2024.101310 [DOI] [Google Scholar]
Tamminga M, MacKenzie L, & Embick D (2020). The dynamics of variation in individuals. Linguistic Inquiry, 51(4), 929–954. [DOI] [PMC free article] [PubMed] [Google Scholar]
Tang C, Hamilton LS, & Chang EF (2017). Intonational speech prosody encoding in the human auditory cortex. Science, 357(6353), 797–801. 10.1126/science.aam8577 [DOI] [PMC free article] [PubMed] [Google Scholar]
Theodore RM, & Monto NR (2019). Distributional learning for speech reflects cumulative exposure to a talker’s phonetic distributions. Psychonomic Bulletin & Review, 26(3), 985–992. 10.3758/s13423-018-1551-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
Ullman MT (2004). Contributions of memory circuits to language: The declarative/procedural model. Cognition, 92(1-2), 231–270. 10.1016/j.cognition.2003.10.008 [DOI] [PubMed] [Google Scholar]
Vroomen J, & Baart M (2009). Phonetic recalibration in audiovisual speech. In Murray MM & Wallace MT (Eds.), The Neural Bases of Multisensory Processes (pp. 363–379). CRC Press. [Google Scholar]
Wade T, & Holt LL (2005). Perceptual effects of preceding nonspeech rate on temporal properties of speech categories. Perception & Psychophysics, 67(5), 939–950. [DOI] [PubMed] [Google Scholar]
Wagner M, & Watson DG (2010). Experimental and theoretical advances in prosody: A review. Language and Cognitive Processes, 25(7-9), 905–945. 10.1080/01690961003589492 [DOI] [PMC free article] [PubMed] [Google Scholar]
Ward G, & Hirschberg J (1985). Implicating uncertainty: The pragmatics of fall-rise intonation. Language, 61(4), 747–776. [Google Scholar]
Warren P (2017). The interpretation of prosodic variability in the context of accompanying sociophonetic cues. Laboratory Phonology, 8(1), 11. 10.5334/labphon.20 [DOI] [Google Scholar]
Weatherholtz K (2015). Perceptual learning of systemic cross-category vowel variation [Doctoral dissertation, Ohio State University; ]. [Google Scholar]
Witteman MJ, Weber A, & McQueen JM (2015). Foreign accent strength and listener familiarity with an accent codetermine speed of perceptual adaptation. Attention, Perception, & Psychophysics, 77(2), 537–556. 10.3758/s13414-014-0785-5 [DOI] [PubMed] [Google Scholar]
Xie X, Earle FS, & Myers EB (2016). Sleep facilitates talker generalization of accent adaptation. The Journal of the Acoustical Society of America, 140(4), 3342. 10.1121/1.4970675 [DOI] [Google Scholar]
Xie X, Earle FS, & Myers EB (2018). Sleep facilitates generalization of accent adaptation to a new talker. Language, Cognition and Neuroscience, 33(2), 196–210. 10.1080/23273798.2017.1369551 [DOI] [PMC free article] [PubMed] [Google Scholar]
Xie X, Buxó-Lugo A, & Kurumada C (2021). Encoding and decoding of meaning through structured variability in speech prosody. Cognition, 211, 104619. 10.1016/j.cognition.2021.104619 [DOI] [PubMed] [Google Scholar]
Xie X, Jaeger TF, & Kurumada C (2023). What we do (not) know about the mechanisms underlying adaptive speech perception: A computational review. Cortex, 166, 377–424. 10.1016/j.cortex.2023.05.003 [DOI] [PubMed] [Google Scholar]
Zhang X, & Samuel AG (2014). Perceptual learning of speech under optimal and adverse conditions. Journal of Experimental Psychology: Human Perception and Performance, 40(1), 200–217. 10.1037/a0033182 [DOI] [PMC free article] [PubMed] [Google Scholar]
Zheng Y, & Samuel AG (2023). Flexibility and stability of speech sounds: The time course of lexically-driven recalibration. Journal of Phonetics, 97, 101222. 10.1016/j.wocn.2023.101222 [DOI] [Google Scholar]

[R1] Anderson MC, & Hulbert JC (2021). Active forgetting: Adaptation of memory by prefrontal control. Annual Review of Psychology, 72, 1–36. 10.1146/annurev-psych-072720-094140 [DOI] [PubMed] [Google Scholar]

[R2] Apfelbaum KS, & McMurray B (2015). Relative cue encoding in the context of sophisticated models of categorization: Separating information from categorization. Psychonomic Bulletin & Review, 22(3), 916–943. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R3] Apfelbaum KS, Kutlu E, McMurray B, Kapnoula EC. (2022). Don’t force it! Gradient speech categorization calls for continuous categorization tasks. J Acoust Soc Am. 152 (6):3728. doi: 10.1121/10.0015201. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R4] Arnhold A, & Kyröläinen A-J (2017). Modelling the interplay of multiple cues in prosodic focus marking. Laboratory Phonology, 8(1). 10.5334/labphon.78 [DOI] [Google Scholar]

[R5] Arnold JE, & Watson DG (2015). Synthesizing meaning and processing approaches to prosody: Performance matters. Language, Cognition and Neuroscience, 30(1-2), 88–102. 10.1080/01690965.2013.840733 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R6] Arthur SE, Wise RJS, Thomas DL, & Leech R (2021). The neural dynamics underlying continuous speech comprehension in naturalistic conditions. Communications Biology, 4(1), 1094. 10.1038/s42003-021-02618-z34535759 [DOI] [Google Scholar]

[R7] Arvaniti A (2019). Crosslinguistic variation, phonetic variability, and the formation of categories in intonation. In Calhoun S, Escudero P, Tabain M, & Warren P (Eds.), Proceedings of the 19th International Congress of Phonetic Sciences, Melbourne, Australia 2019 (pp. 1–6). [Google Scholar]

[R8] Arvaniti A, & Garding G (2007). Dialectal variation in the rising accents of American English. In Cole J & Hualde JI (Eds.), Laboratory Phonology 9 (pp. 547–576). Mouton de Gruyter. [Google Scholar]

[R9] Arvaniti A, Tilsen S, Cangemi F, & Shue Y-L (2024). Continuous and categorical aspects of intonational categories. Journal of Phonetics, 102, 101287. 10.1016/j.wocn.2023.101287 [DOI] [Google Scholar]

[R10] Aylett M, & Turk A (2004). The smooth signal redundancy hypothesis: A functional explanation for relationships between redundancy, prosodic prominence, and duration in spontaneous speech. Language and Speech, 47(1), 31–56. 10.1177/00238309040470010201 [DOI] [PubMed] [Google Scholar]

[R11] Bartle A, & Dellwo V (2015). Rhythmic variability and a repair effect on phonological development. In Proceedings of the 18th International Congress of Phonetic Sciences. University of Glasgow. [Google Scholar]

[R12] Baumann S, & Grice M (2006). The intonation of accessibility. Journal of Pragmatics, 38(10), 1636–1657. 10.1016/j.pragma.2005.03.017 [DOI] [Google Scholar]

[R13] Bent T, & Baese-Berk MM (2021). Listening to non-native speech: Examining the roles of linguistic and indexical information. The Journal of the Acoustical Society of America, 149(2), 1369–1379. [Google Scholar]

[R14] Boersma P, & Weenink D (2012). Praat: Doing phonetics by computer [Computer program]. Version 5.3.23. http://www.praat.org/

[R15] Bolinger D (1986). Intonation and its parts: Melody in spoken English. Stanford University Press. [Google Scholar]

[R16] Bosker HR (2021). Accounting for L1 and L2 listening: Investigating the time-course of speaker-specific processing with reaction times. Language Learning, 71(S1), 38–70. [Google Scholar]

[R17] Bosker HR (2022). Prosodic processing in a second language does not differ between information structure-based and acoustic-phonetic based learners. Bilingualism: Language and Cognition, 25(4), 689–700. 10.1017/S1366728922000281 [DOI] [Google Scholar]

[R18] Bradlow AR, Bassard AM, & Paller KA (2023). Generalized perceptual adaptation to second-language speech: Variability, similarity, and intelligibility. The Journal of the Acoustical Society of America, 154(3), 1601–1613. 10.1121/10.0020914 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R19] Brown-Schmidt S, Yoon SO, & Ryskin RA (2015). People as contexts in conversation. Psychology of Learning and Motivation, 62, 59–99. 10.1016/bs.plm.2014.09.003 [DOI] [Google Scholar]

[R20] Buxó-Lugo A, Toscano JC, & Watson DG (2018). Effects of participant engagement on prosodic prominence. Discourse Processes, 55(3), 305–323. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R21] Calhoun S, & White J (2025). Modeling prosodic prominence in second language acquisition. Language Learning, 75(1), 1–34. 10.1111/lang.12619 [DOI] [Google Scholar]

[R22] Cangemi F, & Grice M (2016). The importance of a distributional approach to categoriality in autosegmental-metrical accounts of intonation. Laboratory Phonology, 7(1). 10.5334/labphon.28 [DOI] [Google Scholar]

[R23] Cangemi F, Krüger M, & Grice M (2015). Listener-specific perception of speaker-specific productions in intonation. In Individual Differences in Speech Production and Perception (pp. 123–145). Peter Lang. [Google Scholar]

[R24] Carvalho PF, Chen CH, & Yu C (2021). The distributional properties of exemplars affect category learning and generalization. Scientific Reports, 11(1), 1–10. 10.1038/s41598-021-90743-0 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R25] Chodroff E, & Cole JS (2019). Testing the distinctiveness of intonational tunes: Evidence from imitative productions in American English. In Interspeech 2019 (pp. 3109–3113). 10.21437/interspeech.2019-2684 [DOI] [Google Scholar]

[R26] Chodroff E, & Wilson C (2020). Perceptual normalization of voice pitch depends on between-talker variability. Attention, Perception, & Psychophysics, 82(7), 3294–3313. [Google Scholar]

[R27] Clarke-Davidson CM, Luce PA, & Sawusch JR (2008). Does perceptual learning in speech reflect changes in phonetic category representation or decision bias? Perception & Psychophysics, 70(4), 604–618. [DOI] [PubMed] [Google Scholar]

[R28] Clayards M, Tanenhaus MK, Aslin RN, & Jacobs RA (2008). Perception of speech reflects optimal use of probabilistic speech cues. Cognition, 108(3), 804–809. 10.1016/j.cognition.2008.04.004 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] Clopper CG, & Smiljanic R (2011). Effects of gender and regional dialect on prosodic patterns in American English. Journal of Phonetics, 39(2), 237–245. 10.1016/j.wocn.2011.02.006 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R30] Cole J (2015). Prosody in context: A review. Language, Cognition and Neuroscience, 30(1-2), 1–31. 10.1080/23273798.2014.963130 [DOI] [Google Scholar]

[R31] Cummings SN, Hodges E, & Theodore RM (in press). Cumulative input sensitivity predicts both attenuation and stability of lexically guided perceptual learning. Psychonomic Bulletin & Review. [DOI] [PubMed] [Google Scholar]

[R32] Cummings SN (2025). Linking lexically guided perceptual learning to statistical patterns in speech input [Doctoral dissertation, University of Connecticut]. [Google Scholar]

[R33] Cummings SN, & Theodore RM (2022). Perceptual learning of multiple talkers: Determinants, characteristics, and limitations. Attention, Perception, & Psychophysics, 84(7), 2335–2359. 10.3758/s13414-022-02556-6 [DOI] [PubMed] [Google Scholar]

[R34] Cutler A, & Swinney DA (1987). Prosody and the development of comprehension. Journal of Child Language, 14(1), 145–167. 10.1017/S0305000900012782 [DOI] [PubMed] [Google Scholar]

[R35] Cychosz M, & Newman RS (2023). Perceptual normalization for speaking rate occurs below the level of the syllable. The Journal of the Acoustical Society of America, 153(3), 1486. 10.1121/10.0017360 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R36] DeKeyser R (2020). Skill acquisition theory. In VanPatten B, Keating GD, & Wulff S (Eds.), Theories in Second Language Acquisition: An Introduction (3rd ed., pp. 83–104). Routledge. [Google Scholar]

[R37] Dellwo V, Leemann A, & Kolly M-J (2015). Rhythmic variability between speakers: Articulatory, prosodic, and linguistic factors. The Journal of the Acoustical Society of America, 137(3), 1513–1528. 10.1121/1.4906837 [DOI] [PubMed] [Google Scholar]

[R38] Drouin JR, & Theodore RM (2019). Lexically guided perceptual learning is robust to task-based changes in listening strategy. The Journal of the Acoustical Society of America, 144(2), 1089. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R39] Earle FS, & Qi Z (2022). Understanding the role of sleep in memory consolidation: A focus on speech perception. Brain and Language, 228, 105095. 10.1016/j.bandl.2022.105095 [DOI] [PubMed] [Google Scholar]

[R40] Earle FS, & Ullman MT (2021). Deficits of learning in procedural memory and consolidation in declarative memory in adults with developmental language disorder. Journal of Speech, Language, and Hearing Research, 64(2), 531–541. 10.1044/2020_JSLHR-20-00292 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R41] Eisner F, & McQueen JM (2006). Perceptual learning in speech: Stability over time. The Journal of the Acoustical Society of America, 119(4), 1950–1953. 10.1121/1.2178721 [DOI] [PubMed] [Google Scholar]

[R42] Fenn KM, Nusbaum HC, & Margoliash D (2003). Consolidation during sleep of perceptual learning of spoken language. Nature, 425(6958), 614–616. 10.1038/nature01951 [DOI] [PubMed] [Google Scholar]

[R43] Gnanateja GN, Grandhi SA, & Cole J (2025). Context-dependent prosodic variation in speech production and perception. Laboratory Phonology, 16(1). 10.16995/labphon.10067 [DOI] [Google Scholar]

[R44] Goldinger SD (1998). Echoes of echoes? An episodic theory of lexical access. Psychological Review, 105(2), 251–279. [DOI] [PubMed] [Google Scholar]

[R45] Goldrick M, & Cole J (2023). Advancement of phonetics in the 21st century: Exemplar models of speech production. Journal of Phonetics, 99, 101254. 10.1016/j.wocn.2023.101254 [DOI] [Google Scholar]

[R46] Goudbeek M, Swingley D, & Smits R (2009). Supervised and unsupervised learning of multidimensional acoustic categories. Journal of Experimental Psychology: Human Perception and Performance, 35(6), 1913–1933. 10.1037/a0015781 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R47] Grice M, Ritter S, Niemann H, & Roettger TB (2017). Integrating the discreteness and continuity of intonational categories. Journal of Phonetics, 64, 90–107. 10.1016/j.wocn.2017.03.003 [DOI] [Google Scholar]

[R48] Gryllia S, Baltazani M, & Arvaniti A (2018). The role of pragmatics and politeness in explaining prosodic variability. In Proceedings of the 9th International Conference on Speech Prosody 2018 (pp. 32–36). [Google Scholar]

[R49] Gunlogson C (2008). A question of commitment. Belgian Journal of Linguistics, 22(1), 101–136. 10.1075/bjl.22.06gun [DOI] [Google Scholar]

[R50] Hansen M, Horne M, Lindström E, & Roll M (2015). Neurocognitive processes of prosodic sentence type recognition in L2. In Babatsouli E & Ingram D (Eds.), Proceedings of the International Symposium on Monolingual and Bilingual Speech 2015 (pp. 174–181). [Google Scholar]

[R51] Heffner CC, & Myers EB (2021). Individual differences in phonetic plasticity across native and nonnative contexts. Journal of Speech, Language, and Hearing Research, 64(9), 3720–3733. 10.1044/2021_JSLHR-20-00650 [DOI] [PubMed] [Google Scholar]

[R52] Heffner CC, Fuhrmeister P, Luthra S, Mechtenberg H, Saltzman D, & Myers EB (2022). Reliability and validity for perceptual flexibility in speech. Brain and Language, 226, 105070. 10.1016/j.bandl.2021.105070 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R53] Hitczenko K, & Feldman NH (2022). Naturalistic speech supports distributional learning across contexts. Proceedings of the National Academy of Sciences, 119(38), e2123230119. 10.1073/pnas.2123230119 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R54] Holliday JJ (2021). Context effects on vowel-to-vowel coarticulation in three dialects of American English. Language and Speech, 64(3), 694–720. 10.1177/0023830920951895 [DOI] [Google Scholar]

[R55] Huttenlauch C, De Beer C, Hanne S, & Wartenburger I (2021). Production of prosodic cues in coordinate name sequences addressing varying interlocutors. Laboratory Phonology: Journal of the Association for Laboratory Phonology, 12(1), 1. 10.5334/labphon.221 [DOI] [Google Scholar]

[R56] Idemaru K, & Holt LL (2011). Word recognition reflects dimension-based statistical learning. Journal of Experimental Psychology: Human Perception and Performance, 37(6), 1939–1956. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R57] Idemaru K, & Holt LL (2020). Talker-contingent auditory perceptual learning generalizes across voices and depends on early sensory processing. Attention, Perception, & Psychophysics, 82(1), 191–210. [Google Scholar]

[R58] Jesse A, & Kaplan E (2019). Reward sensitivity and perceptual learning in speech. Attention, Perception, & Psychophysics, 81(5), 1367–1384. 10.3758/s13414-018-01651-x [DOI] [PubMed] [Google Scholar]

[R59] Jessen M (2008). Forensic phonetics. Language and Linguistics Compass, 2(4), 671–711. 10.1111/j.1749-818X.2008.00066.x [DOI] [Google Scholar]

[R60] Johnson K (1997). Speech perception without speaker normalization: An exemplar model. In Johnson K & Mullennix JW (Eds.), Talker Variability in Speech Processing (pp. 145–165). Academic Press. [Google Scholar]

[R61] Johnson K, & Sjerps MJ (2021). Speaker normalization in speech perception. In Pardo JS, Nygaard LC, Remez RE, & Pisoni DB (Eds.), The Handbook of Speech Perception (2nd ed., pp. 145–176). Wiley. 10.1002/9781119184096.ch6 [DOI] [Google Scholar]

[R62] Jun S-A, & Bishop J (2015). Priming implicit prosody: Prosodic boundaries and individual differences. Language and Speech, 58(4), 459–473. 10.1177/0023830914563368 [DOI] [PubMed] [Google Scholar]

[R63] Kaschak MP, Kutta TJ, & Jones JL (2011). Structural priming as implicit learning: Cumulative priming effects and individual differences. Psychonomic Bulletin & Review, 18(6), 1133–1139. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R64] Kleinschmidt DF, & Jaeger TF (2015). Robust speech perception: Recognize the familiar, generalize to the similar, and adapt to the novel. Psychological Review, 122(2), 148–203. 10.1037/a0038695 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R65] Kronrod Y, Coppess E, & Feldman NH (2016). A unified model of categorical effects in consonant and vowel perception. Psychonomic Bulletin & Review, 23(6), 1681–1712. 10.3758/s13423-016-1049-y [DOI] [PubMed] [Google Scholar]

[R66] Kuhl PK, Conboy BT, Padden D, Nelson T, & Pruitt J (2010). Early speech perception and later language development: Implications for the critical period. Language Learning and Development, 1(4), 237–264. [Google Scholar]

[R67] Kurumada C, & Buxó-Lugo A (2024). Intonation adaptation to multiple talkers. Journal of Experimental Psychology: Learning, Memory, and Cognition, 50(12), 1954–1981. 10.1037/xlm0001419 [DOI] [PubMed] [Google Scholar]

[R68] Kurumada C, & Roettger TB (2022). Thinking probabilistically in the study of intonational speech prosody. WIREs Cognitive Science, 13(1), e1579. 10.1002/wcs.1579 [DOI] [PubMed] [Google Scholar]

[R69] Kurumada C, Xie X, & Buxó-Lugo A (2024). Talker-specific adaptation in prosodic processing: Investigating rapid perceptual adjustment to interlocutor variation. Cognition, 242, 105638.37839251 [Google Scholar]

[R70] Ladd DR (2008). Intonational Phonology (2nd ed.). Cambridge University Press. [Google Scholar]

[R71] Ladd DR, & Morton R (1997). The perception of intonational emphasis: Continuous or categorical? Journal of Phonetics, 25(3), 313–342. 10.1006/S0095-4470(97)90063-X [DOI] [Google Scholar]

[R72] Lavan N, Burton AM, Scott SK, & McGettigan C (2024). Voice identity processing in unfamiliar, newly learned, and familiar voices: A systematic review. Psychonomic Bulletin & Review, 31(1), 67–92. 10.3758/s13423-023-02333-9 [DOI] [Google Scholar]

[R73] Lee C-Y, Yen H-L, Yeh P-W, Lin W-H, Cheng Y-Y, Tzeng Y-L, & Wu H-C (2010). Fluent speech language representation in the brain: Effects of frequency and word class. Neuropsychologia, 48(10), 2988–3000. 10.1016/j.neuropsychologia.2010.06.008 [DOI] [Google Scholar]

[R74] Li J, Luo J, & Yu L (2021). Talker normalization of vowel categories in auditory cortex: Evidence from fMRI-based multivariate pattern analyses. The Journal of the Acoustical Society of America, 149(4), 2643–2654. [Google Scholar]

[R75] Liu L, & Holt LL (2015). Dimension-based statistical learning of vowels. Journal of Experimental Psychology: Human Perception and Performance, 41(6), 1783–1798. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R76] Liu L, & Jaeger TF (2018). Inferring causes during speech perception. Cognition, 174, 55–70. 10.1016/j.cognition.2018.01.003 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R77] Liu L, & Jaeger TF (2019). Talker-specific effects on spoken word recognition in immediate and delayed memory tests. Attention, Perception, & Psychophysics, 81(7), 2215–2233. [Google Scholar]

[R78] Lobanov BM (1971). Classification of Russian vowels spoken by different speakers. The Journal of the Acoustical Society of America, 49(2B), 606–608. [Google Scholar]

[R79] Lorenzen J, & Baumann S (2024). Adaptive prosody processing: Effects of prosody type, processing load, and adaptation direction. Journal of Phonetics, 103, 101311. 10.1016/j.wocn.2024.101311 [DOI] [Google Scholar]

[R80] Luo Y, Mitterer H, Zhou X, & Chen Y (2024). Flexibility and stability in lexical tone recalibration: Evidence from tone perceptual learning. Language and Speech, 00238309241291536. 10.1177/00238309241291536 [DOI] [PubMed] [Google Scholar]

[R81] Luthra S, Guediche S, Blumstein SE, & Myers EB (2021). Neural substrates of subphonemic variation and lexical competition in spoken word recognition. Language, Cognition and Neuroscience, 36(2), 151–169. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R82] Maddox WT, Chandrasekaran B, Smayda K, & Yi H-G (2013). Dual systems of speech category learning across the lifespan. Psychology and Aging, 28(4), 1042–1056. 10.1037/a0034969 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R83] Malisz Z, Bröker F, & Andreeva B (2018). Dimensions of rhythmic variability in spontaneous speech. In Proceedings of the 9th International Conference on Speech Prosody 2018 (pp. 1054–1058). 10.21437/SpeechProsody.2018-213 [DOI] [Google Scholar]

[R84] Maslowski M, Meyer AS, & Bosker HR (2019). How the tracking of habitual rate influences speech perception. Journal of Experimental Psychology: Learning, Memory, and Cognition, 45(1), 128–138. [DOI] [PubMed] [Google Scholar]

[R85] McAuliffe MJ, & Babel M (2016). Stimulus-directed attention attenuates lexically-guided perceptual learning. The Journal of the Acoustical Society of America, 140(3), 1727–1738. [DOI] [PubMed] [Google Scholar]

[R86] McMurray B, & Jongman A (2011). What information is necessary for speech categorization? Harnessing variability in the speech signal by integrating cues computed relative to expectations. Psychological Review, 118(2), 219–246. 10.1037/a0022325 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R87] McQueen JM, & Dilley LC (2020). Prosody and spoken-word recognition. In Gussenhoven C & Chen A (Eds.), The Oxford Handbook of Language Prosody (pp. 554–565). Oxford University Press. 10.1093/oxfordhb/9780198832232.013.33 [DOI] [Google Scholar]

[R88] Mechtenberg H, Cheng L, Luthra S, & Myers EB (2025). Cognitive load does not affect lexically-guided perceptual learning. Attention, Perception, & Psychophysics. [Google Scholar]

[R89] Milne AE, Petkov CI, & Wilson B (2021). Auditory and visual sequence learning in humans and monkeys using an artificial grammar learning paradigm. Neuroscience, 389, 205–217. 10.1016/j.neuroscience.2017.06.059 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R90] Myers EB, & Mesite LM (2014). Neural systems underlying perceptual adjustment to non-standard speech tokens. Journal of Memory and Language, 76, 80–93. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R91] Nakamura C, White L, & Kondo M (2021). Rate normalization in the perception of L2 prosody. The Journal of the Acoustical Society of America, 150(4), A82–A82. [Google Scholar]

[R92] Nearey TM (1989). Static, dynamic, and relational properties in vowel perception. The Journal of the Acoustical Society of America, 85(5), 2088–2113. [DOI] [PubMed] [Google Scholar]

[R93] Newman RS, & Sawusch JR (2009). Perceptual normalization for speaking rate III: Effects of the rate of one voice on perception of another. Journal of Phonetics, 37(1), 46–65. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R94] Norris D, & McQueen JM (2008). Shortlist B: A Bayesian model of continuous speech recognition. Psychological Review, 115(2), 357–395. 10.1037/0033-295X.115.2.357 [DOI] [PubMed] [Google Scholar]

[R95] Norris D, McQueen JM, & Cutler A (2003). Perceptual learning in speech. Cognitive Psychology, 47(2), 204–238. [DOI] [PubMed] [Google Scholar]

[R96] Nota N, Trujillo J, & De Ruiter JP (2023). Variability in gesture production and perception. Language and Cognition, 15(2), 241–268. 10.1017/langcog.2022.31 [DOI] [Google Scholar]

[R97] Nusbaum HC, & Magnuson JS (1997). Talker normalization: Phonetic constancy as a cognitive process. In Johnson K & Mullennix JW (Eds.), Talker Variability in Speech Processing (pp. 109–132). Academic Press. [Google Scholar]

[R98] Nygaard LC, & Pisoni DB (1998). Talker-specific learning in speech perception. Perception & Psychophysics, 60(3), 355–376. 10.3758/BF03206860 [DOI] [PubMed] [Google Scholar]

[R99] Ostrow A, & Kurumada C (2025). Short-term and long-term adaptive changes in prosodic comprehension. In Proceedings of the 38th Annual Conference on Human Sentence Processing. University of Michigan. [Google Scholar]

[R100] Ostrow A & Kurumada C, (2025, April 18). Short-term and long-term adaptive changes in prosodic comprehension. OSF. https://osf.io/3xvsh/ [DOI] [PMC free article] [PubMed]

[R101] Patel R, Connaghan K, Franco D, Edsall E, Forgit D, Olsen L, & Russell S (2011). The Caterpillar: A novel reading passage for assessment of motor speech disorders. American Journal of Speech-Language Pathology, 22(1), 1–9. [DOI] [PubMed] [Google Scholar]

[R102] Persson F, & Jaeger TF (2023). Continuous adaptation in native and non-native speech perception: Implications for theories of adaptation. The Journal of the Acoustical Society of America, 153(6), 3368–3380. [Google Scholar]

[R103] Pickering MJ, & Branigan HP (1998). The representation of verbs: Evidence from syntactic priming in language production. Journal of Memory and Language, 39(4), 633–651. 10.1006/jmla.1998.2592 [DOI] [Google Scholar]

[R104] Pierrehumbert JB (2001). Exemplar dynamics: Word frequency, lenition and contrast. In Bybee J & Hopper P (Eds.), Frequency and the Emergence of Linguistic Structure (pp. 137–157). John Benjamins. [Google Scholar]

[R105] Pierrehumbert JB, & Hirschberg J (1990). The meaning of intonational contours in the interpretation of discourse. In Cohen PR, Morgan J, & Pollack ME (Eds.), Intentions in Communication (pp. 271–311). MIT Press. [Google Scholar]

[R106] Pisanski K, Cartei V, McGettigan C, Raine J, & Reby D (2017). Voice modulation: A window into the origins of human vocal control? Trends in Cognitive Sciences, 20(4), 304–318. 10.1016/j.tics.2016.01.002 [DOI] [PubMed] [Google Scholar]

[R107] Podesva RJ, & Callier P (2015). Voice quality and identity. Annual Review of Applied Linguistics, 35, 173–194. 10.1017/S0267190514000270 [DOI] [Google Scholar]

[R108] Quam C, & Creel SC (2021). Impacts of acoustic-phonetic variability on perceptual development for spoken language: A review. Wiley Interdisciplinary Reviews: Cognitive Science, 12(3), e1558. 10.1002/wcs.1558 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R109] Quam C, Wang A, Maddox WT, Golisch K, & Lotto A (2018). Procedural-memory, working-memory, and declarative-memory skills are each associated with dimensional integration in sound-category learning. Frontiers in Psychology, 9, 1828. 10.3389/fpsyg.2018.01828 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R110] R Core Team. (2024). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://www.R-project.org/s [Google Scholar]

[R111] Roettger TB, & Franke M (2019). Evidential strength of intonational cues and rational adaptation to (un-)reliable intonation. Cognitive Science, 43(7), e12745. 10.1111/cogs.12745 [DOI] [PubMed] [Google Scholar]

[R112] Roettger TB, & Rimland K (2020). Listeners rapidly adapt to new talker prosody: Evidence from adaptation priming. In Proceedings of the 10th International Conference on Speech Prosody 2020 (pp. 844–848). 10.21437/SpeechProsody.2020-172 [DOI] [Google Scholar]

[R113] Saltzman D, & Myers EB (2021). The time course of individual learning history effects on perceptual adjustment. Attention, Perception, & Psychophysics, 83(3), 989–1009. 10.3758/s13414-020-02150-x [DOI] [Google Scholar]

[R114] Samuel AG, & Dumay N (2021). Auditory selective adaptation moment by moment, at multiple timescales. Journal of Experimental Psychology: Human Perception and Performance, 47(4), 596–615. 10.1037/xhp0000841 [DOI] [PubMed] [Google Scholar]

[R115] Samuel AG, & Kraljic T (2009). Perceptual learning for speech. Attention, Perception, & Psychophysics, 71(6), 1207–1218. 10.3758/APP.71.6.1207 [DOI] [PubMed] [Google Scholar]

[R116] Severijnen GGA, Bosker HR, Plag I, & McQueen JM (2021). Listeners track talker-specific prosody to deal with talker-variability. Brain Research, 1755, 147477. 10.1016/j.brainres.2020.147477 [DOI] [PubMed] [Google Scholar]

[R117] Severijnen GGA, Di Dona G, Bosker HR, & McQueen JM (2023). Tracking talker-specific cues to lexical stress: Evidence from perceptual learning. Journal of Experimental Psychology: Human Perception and Performance, 49(4), 549–565. 10.1037/xhp0001105 [DOI] [PubMed] [Google Scholar]

[R118] Severijnen GGA, Bosker HR, & McQueen JM (2024a). Learning to use talker-specific cues to lexical stress: More rapidly when staying with the same talker. Attention, Perception, & Psychophysics, 86(1), 63–81. 10.3758/s13414-023-02789-3 [DOI] [Google Scholar]

[R119] Severijnen GGA, Bosker HR, & McQueen JM (2024b). Your “VOORnaam” is not my “VOORnaam”: An acoustic analysis of individual talker differences in word stress in Dutch. Journal of Phonetics 103, 101296. doi: 10.1016/j.wocn.2024.101296 [DOI] [Google Scholar]

[R120] Severijnen GGA, McQueen JM, & Bosker HR (2025). Cross-talker generalization of talker-specific phonetic variation in speech perception. Journal of Phonetics, 98, 101243. [Google Scholar]

[R121] Sharf DJ, & Ohde RN (1981). Physiologic, acoustic, and perceptual aspects of coarticulation: Implications for the remediation of articulatory disorders. In Lass NJ (Ed.), Speech and Language: Advances in Basic Research and Practice (Vol. 5, pp. 153–247). Academic Press. [Google Scholar]

[R122] Sjerps MJ, Fox NP, Johnson K, & Chang EF (2019). Speaker-normalized sound representations in the human auditory cortex. Nature Communications, 10(1), 2465. 10.1038/s41467-019-10365-z [DOI] [PMC free article] [PubMed] [Google Scholar]

[R123] Steffman J, & Cole J (2024). Acoustic and perceptual evidence for systematic variation in English intonation. Phonology, 41(1), 1–35. 10.1017/S0952675723000271 [DOI] [Google Scholar]

[R124] Steffman J, Cole J, & Shattuck-Hufnagel S (2024). Intonational categories and continua in American English rising nuclear tunes. Journal of Phonetics, 104, 101310. 10.1016/j.wocn.2024.101310 [DOI] [Google Scholar]

[R125] Tamminga M, MacKenzie L, & Embick D (2020). The dynamics of variation in individuals. Linguistic Inquiry, 51(4), 929–954. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R126] Tang C, Hamilton LS, & Chang EF (2017). Intonational speech prosody encoding in the human auditory cortex. Science, 357(6353), 797–801. 10.1126/science.aam8577 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R127] Theodore RM, & Monto NR (2019). Distributional learning for speech reflects cumulative exposure to a talker’s phonetic distributions. Psychonomic Bulletin & Review, 26(3), 985–992. 10.3758/s13423-018-1551-5 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R128] Ullman MT (2004). Contributions of memory circuits to language: The declarative/procedural model. Cognition, 92(1-2), 231–270. 10.1016/j.cognition.2003.10.008 [DOI] [PubMed] [Google Scholar]

[R129] Vroomen J, & Baart M (2009). Phonetic recalibration in audiovisual speech. In Murray MM & Wallace MT (Eds.), The Neural Bases of Multisensory Processes (pp. 363–379). CRC Press. [Google Scholar]

[R130] Wade T, & Holt LL (2005). Perceptual effects of preceding nonspeech rate on temporal properties of speech categories. Perception & Psychophysics, 67(5), 939–950. [DOI] [PubMed] [Google Scholar]

[R131] Wagner M, & Watson DG (2010). Experimental and theoretical advances in prosody: A review. Language and Cognitive Processes, 25(7-9), 905–945. 10.1080/01690961003589492 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R132] Ward G, & Hirschberg J (1985). Implicating uncertainty: The pragmatics of fall-rise intonation. Language, 61(4), 747–776. [Google Scholar]

[R133] Warren P (2017). The interpretation of prosodic variability in the context of accompanying sociophonetic cues. Laboratory Phonology, 8(1), 11. 10.5334/labphon.20 [DOI] [Google Scholar]

[R134] Weatherholtz K (2015). Perceptual learning of systemic cross-category vowel variation [Doctoral dissertation, Ohio State University; ]. [Google Scholar]

[R135] Witteman MJ, Weber A, & McQueen JM (2015). Foreign accent strength and listener familiarity with an accent codetermine speed of perceptual adaptation. Attention, Perception, & Psychophysics, 77(2), 537–556. 10.3758/s13414-014-0785-5 [DOI] [PubMed] [Google Scholar]

[R136] Xie X, Earle FS, & Myers EB (2016). Sleep facilitates talker generalization of accent adaptation. The Journal of the Acoustical Society of America, 140(4), 3342. 10.1121/1.4970675 [DOI] [Google Scholar]

[R137] Xie X, Earle FS, & Myers EB (2018). Sleep facilitates generalization of accent adaptation to a new talker. Language, Cognition and Neuroscience, 33(2), 196–210. 10.1080/23273798.2017.1369551 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R138] Xie X, Buxó-Lugo A, & Kurumada C (2021). Encoding and decoding of meaning through structured variability in speech prosody. Cognition, 211, 104619. 10.1016/j.cognition.2021.104619 [DOI] [PubMed] [Google Scholar]

[R139] Xie X, Jaeger TF, & Kurumada C (2023). What we do (not) know about the mechanisms underlying adaptive speech perception: A computational review. Cortex, 166, 377–424. 10.1016/j.cortex.2023.05.003 [DOI] [PubMed] [Google Scholar]

[R140] Zhang X, & Samuel AG (2014). Perceptual learning of speech under optimal and adverse conditions. Journal of Experimental Psychology: Human Perception and Performance, 40(1), 200–217. 10.1037/a0033182 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R141] Zheng Y, & Samuel AG (2023). Flexibility and stability of speech sounds: The time course of lexically-driven recalibration. Journal of Phonetics, 97, 101222. 10.1016/j.wocn.2023.101222 [DOI] [Google Scholar]

PERMALINK

Short-term and long-term adaptive changes in prosodic comprehension

Amelia Ostrow

Chigusa Kurumada

Abstract

1. Introduction

1.1. How do listeners navigate the prosodic variability?

1.2. Stability of cross-talker adaptation

1.3. The current study

2. Methods

2.1. Participants

Table 1.

2.2. Stimuli

2.2.1. Adaptation task stimuli

Figure 1.

2.2.2. Discrimination task stimuli

2.3. Procedure

Figure 2.

2.4. Predictions

Prediction 1: Listeners adapt their prosodic categorization within each session (i.e., one day)

Prediction 2: Listeners maintain their prosodic adaptation for two sessions

Prediction 3: The maintenance of the adaptation decays over time

3. Results

3.1. Discrimination task

Figure 3.

3.2. Social inference questionnaire

Figure 4.

3.3. Adaptation task

Figure 5.

Figure 6.

3.3.1. Question-biasing condition

Table 2.

3.3.2. Statement-biasing condition

Table 3.

3.3.3. Question- and statement-biasing conditions comparison

3.4. Exposure phase responses

Figure 7.

4. General discussion

4.1. Longevity of intonation adaptation

4.2. Memory encoding of intonation in the moment and over time

4.3. Did participants use task-specific decision rules or strategies?

4.4. Limitations of the current study and future directions

5. Appendix

Appendix I. Participant Demographics

Table A1.

Table A2.

Figure A1.

Figure A2.

Appendix II. Mixed effect logistic regression models for the discrimination task

Table A3.

Figure A3.

Appendix III. Question- and statement-biasing conditions comparison

Table A4.

Appendix IV. Response accuracy and reaction times in the exposure phase

Figure A4.

Appendix V. Mixed effect logistic regression for exposure phase responses

Table A5.

Table A6.

Table A7.

Appendix VI. Exploratory analyses on the inferred age/location and adaptation responses

1. Age

Table A8.

2. Geographical region

Table A9.

Appendix VII. Comparisons of exposure and test stimulus features

Figure A5.

References

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases