Vowel generalization and its relation to adaptation during perturbations of auditory feedback

Speech adaptations to alterations, or perturbations, of auditory feedback have provided important insights into sensorimotor representations underlying speech. One finding from these studies that is yet to be accounted for is vowel generalization, which describes the effects of repeated perturbations to one vowel on the production of other vowels that were not perturbed. The present study used correlation testing to quantify the effects of changes in a perturbed vowel on neighboring (i.e., similar) nonperturbed vowels. The results identified significant correlations between the changes of adjacent, but not nonadjacent, vowel pairs. This finding suggests that generalization is partly a response to adaptation and not solely due to the auditory perturbation.

Abstract

Repeated perturbations of auditory feedback during vowel production elicit changes not only in the production of the perturbed vowel (adaptation) but also in the production of nearby vowels that were not perturbed (generalization). The finding that adaptation generalizes to other, nonperturbed vowels suggests that sensorimotor representations for vowels are not independent; instead, the goals for producing any one vowel may depend in part on the goals for other vowels. The present study investigated the dependence or independence of vowel representations by evaluating adaptation and generalization in two groups of speakers exposed to auditory perturbations of their first formant (F1) during different vowels. The speakers in both groups who adapted to the perturbation exhibited generalization in two nonperturbed vowels that were produced under masking noise. Correlation testing was performed to evaluate the relations between adaptation and generalization as well as between the generalization in the two nonperturbed vowels. These tests identified significant coupling between the F1 changes of adjacent vowels but not nonadjacent vowels. The pattern of correlation findings indicates that generalization was due in part to feedforward representations that are partly shared across adjacent vowels, possibly to maintain their acoustic contrast.

NEW & NOTEWORTHY Speech adaptations to alterations, or perturbations, of auditory feedback have provided important insights into sensorimotor representations underlying speech. One finding from these studies that is yet to be accounted for is vowel generalization, which describes the effects of repeated perturbations to one vowel on the production of other vowels that were not perturbed. The present study used correlation testing to quantify the effects of changes in a perturbed vowel on neighboring (i.e., similar) nonperturbed vowels. The results identified significant correlations between the changes of adjacent, but not nonadjacent, vowel pairs. This finding suggests that generalization is partly a response to adaptation and not solely due to the auditory perturbation.

the speed at which sounds and syllables are articulated attests to the importance of feedforward control during speech production (Houde and Chang 2015; Tsao and Weismer 1997). At the same time, the finding that speakers rapidly respond to alterations, or perturbations, of auditory (Heinks-Maldonado and Houde 2005; Larson et al. 2000; Reilly and Dougherty 2013; Tourville et al. 2008) and somatosensory (Abbs and Gracco 1984; Kelso et al. 1984) feedback highlights the importance of feedback control to speech production. For example, perturbations to auditory feedback that increase a vowel’s first formant frequency (F1) elicit rapid compensations that decrease the F1 of that vowel (Niziolek and Guenther 2013; Reilly and Dougherty 2013; Tourville et al. 2008). With repeated exposure, speakers adapt to the auditory perturbation such that compensations become encoded in the feedforward commands for producing the perturbed vowel (Cai et al. 2010; Houde and Jordan 1998; Mitsuya et al. 2011; Villacorta et al. 2007). Several studies have also reported that repeated perturbations to one vowel elicit similar changes in other, nonperturbed vowels (Cai et al. 2010; Houde and Jordan 1998; Rochet-Capellan et al. 2012; Villacorta et al. 2007). This phenomenon, known as generalization, has not been investigated as extensively as either compensation or adaptation, and as a result, the mechanisms responsible for vowel generalization are not as well understood.

Contemporary models of speech production typically characterize vowel targets as regions in an auditory-perceptual space defined by the frequencies of F1, F2, and F3 (Guenther and Perkell 2004; Guenther and Vladusich 2012; Tourville and Guenther 2011). This characterization of vowel targets has accounted for a range of neural and behavioral findings regarding auditory feedback control including compensation and adaptation (Guenther and Hickok 2015; Houde and Chang 2015; Tourville et al. 2008). Despite their success in accounting for many aspects of sensorimotor integration during speech, these models are not currently capable of accounting for generalization (Cai et al. 2010). Houde and Jordan (1998) first described generalization in their investigation of auditory feedback perturbations of vowel productions. These investigators examined responses to repeated auditory perturbations that increased or decreased F1 and F2 during production of the vowel /ɛ/ in consonant vowel consonant (CVC) words such as “pep” and “bep.” The authors reported that repeated auditory perturbation of this vowel elicited compensatory changes in F1 and F2 that were in the direction opposite the perturbation to each formant. In addition, the authors observed that the auditory perturbation elicited similar changes to other vowels such as /i/ (as in “pip”) and /æ/ (as in “pap”) that were not perturbed during the study. During production of these latter vowels, speakers heard a masking noise that blocked their auditory feedback. The generalization of an auditory perturbation to other, nonperturbed vowels has since been replicated in nonwhispered vowel productions (Cai et al. 2010; Rochet-Capellan et al. 2012; Villacorta et al. 2007), as well as diphthong and triphthong vowel productions (Cai et al. 2010).

The difficulty that contemporary models have accounting for generalization may be due to the independence of vowel targets in these models. That is, the finding that changes in the production of one vowel are accompanied by changes in the production of other vowels is difficult to reconcile with models that posit independent vowel targets. Rather, generalization suggests that sensorimotor representations for different vowels are linked in a dependent manner. That generalization results from control parameters linking different vowel targets is supported by recent findings. Specifically, a number of studies have observed that responses to auditory perturbations are influenced by the locations of vowels around the perturbed vowel. For example, Mitsuya and colleagues (Mitsuya et al. 2011, 2013) examined cross-language effects on responses to perturbations of auditory feedback. These investigators observed that perturbation of the same vowel in different languages is associated with greater adaptation when the language has an adjacent vowel in the direction of the perturbation or when an adjacent vowel is closer in one language than another. Moreover, Reilly and Dougherty (2013) and Niziolek and Guenther (2013) observed that auditory perturbations causing large reductions in the distance between a perturbed and neighboring vowel elicit larger compensations than identical perturbations causing smaller reductions in the distance between the two vowels. Research findings demonstrating that the presence and proximity of neighboring vowels modulate responses to auditory perturbations are not consistent with independent vowel representations.

The present study investigated the relationship between generalization and adaptation and whether, or to what extent, generalization resulted from nonindependent vowel representations. Speakers were assigned to two groups and exposed to auditory perturbations of different vowels with different distributions of neighboring vowels. In one analysis, a finding of between-group differences in adaptation would be followed by analyses to determine whether adaptation differences were accompanied by similar differences in generalization. A second analysis evaluated correlations between within-speaker changes associated with adaptation and generalization and between different nonperturbed vowels.

METHODS

Participants.

Forty speakers participated in this study (32 women and 8 men; mean age = 22.9 yr, SD = 5.7 yr). All participants were native speakers of English with no reported history of speech, language, or hearing disorders. On enrolling in the study, speakers were assigned to one of two experimental groups, the pert_æ group or the pert_i group. Data from one speaker in the pert_æ group were excluded from the study because this speaker consistently produced the vowel /æ/ as a diphthong. As a result, the final data set consisted of 20 speakers in the pert_i group (16 women and 4 men; mean age = 21.7, SD = 5.1) and 19 speakers in the pert_æ group (16 women and 3 men; mean age = 23.6, SD = 5.9). Each speaker gave written, informed consent before participating in the study, which was approved by the Northeastern University Institutional Review Board.

Experimental protocol.

The present study evaluated productions of three front vowels lying along consecutive regions of the F1 dimension of the vowel space:/i/ as in “hid,”/ɛ/ as in “head,” and /æ/ as in “had.” The locations of these vowels are illustrated in Fig. 1, which depicts the two-dimensional vowel space of a pilot participant who produced each of the Peterson and Barney (1952) stimuli 10 times. These vowels were embedded in monosyllabic, consonant-vowel-consonant (CVC) words and divided into “training” and “test” words as displayed in Table 1. For one speaker group, the pert_i group, training words all contained the vowel /i/, and for the other group, the pert_æ group, training words contained the vowel /æ/. To control for formant differences between tense (e.g., “bag”) and lax (e.g., “bat”) variants of /æ/ (Gordon 2005), words were limited to those that elicited lax productions of this vowel. The training words comprised six words whose initial and final consonants were, with one exception, identical for each speaker group. During production of these words, speakers received auditory feedback of their speech output, and this feedback was subject to a perturbation of F1 that gradually introduced and then held constant over the course of the experiment (see Experimental design). There was reason to believe that F1 perturbation of /i/ and /æ/ would elicit different patterns of adaptation vowels because of differences in the distributions of their neighboring vowels (see Fig. 1 for reference). Specifically, perturbations that increase F1 in the pert_i group would “push” /i/ toward the adjacent vowel, /ɛ/, and were expected to elicit a larger adaptive response (i.e., decreases in F1) than perturbations of /æ/, which does not have an adjacent vowel along the path of the perturbation (Hillenbrand et al. 1995; Peterson and Barney 1952; Turner et al. 1995; Yang 1996).

Fig. 1.

Standard deviation ellipses depicting vowel locations in a female speaker’s F1, F2 vowel space. Ellipses were based on F1 and F2 values from 10 productions of each of the Peterson and Barney (1952) stimuli.

Table 1.

Training and test words for pert_i and pert_æ speaker groups

Pert_i		Pert_æ
Training words	Test words	Training words	Test words
bit	bat	bat	bit
pit	pat	pat	pit
tick	tech	tack	tech
tip	bet	tap	bet
kit	pick	cat	pack
kid		cap

Test words containing the perturbed vowel have been underlined in each group.

Speakers also produced test words consisting of five CVC words that were not perturbed during the experiment. During production of test words, auditory feedback was replaced with an 87-dB speech-shaped masking noise to minimize speakers’ perception of their speech output. For each speaker group, the test words comprised two words containing the vowel /ɛ/ and two words containing the vowel that was perturbed in the other speaker group (either /i/ or /ae/). As a result, generalization of the F1 perturbation to nonperturbed vowels was evaluated for the vowel /ɛ/ in both groups and for the vowel that was perturbed in the other speaker group. The fifth test word contained the perturbed vowel for each speaker group and was included to examine whether or to what extent adaptive changes in the perturbed vowel were retained when feedback was masked and the vowel was produced in a novel phonetic context (Houde and Jordan 1998; Villacorta et al. 2007).

Experimental setup.

During the experiment, speakers were seated in a sound treated audio booth (model RE-147S; Acoustic Systems) with visual access to a computer monitor that displayed the word for each trial. Words were presented using the Psychophysics Toolbox (Brainard 1997; Kleiner et al. 2007), using custom MATLAB scripts that also controlled the software for perturbing auditory feedback (see below). Speech signals were transduced using a head-worn directional microphone (model C520; AKG) placed at a fixed distance of ~5.5 cm from the speaker’s lips. Microphone signals were amplified using a Mackie VLZ3 mixer/preamp, and left and right microphone channels were digitized by an external sound card (Delta 44; M-Audio; digital sampling rate = 48,000 Hz). Perturbation of vowel F1 was applied to one of the microphone channels and was accomplished using a MATLAB Mex-based software package, Audapter, developed by Cai and colleagues (Cai et al. 2008; Tourville et al. 2013). Audapter estimates the first four formant frequencies by using linear predictive coding (LPC) in combination with cepstral liftering and dynamic programming (Xia and Espy-Wilson 2000). Perturbation of a speaker’s F1 is accomplished by simultaneously zeroing out the input F1 and introducing a pole at the perturbed F1. The order of the LPC analysis in this study was set to 11 for female speakers and 12 for male speakers. Formant tracking and perturbation were initiated and terminated on the basis of two short-term root mean square (rms) measures reflecting rms amplitude of the raw microphone and preemphasized microphone signals. Audio output from Audapter was delivered via the external sound card to a pair of noise-isolating insert earphones (ER4 microPro; Etymotic Research) worn by the speaker. During production of training words, the gain of the feedback signal relative to the microphone input level was ~16 dB SPL and the auditory feedback delay was ~11 ms. After each trial, copies of the microphone signal and the auditory feedback signal were saved to the computer’s hard drive.

Because noise-induced increases in speech intensity can affect formant frequencies (Perkell et al. 2007), Villacorta et al. (2007) used a visual display to help speakers keep their intensity levels within a prescribed range of 67 to 71 dB SPL. A similar approach was used in the present study and involved a virtual sound bar that was rendered on the computer monitor next to the word. Sound pressure level (SPL) was derived from the second microphone channel, which was sampled using sound drivers in Psychophysics Toolbox. The sound bar was present during production of both training and test words, and its height and color provided speakers with continuous information regarding the SPL of their speech output. The height of the vertical bar varied with SPL and was green when the speaker’s SPL was within the target range (67–71 dB SPL), yellow when the speaker’s SPL was below the target range, and red when the speaker’s SPL was above the target range. The refresh rate of the sound bar was ~30 Hz. The display of the sound bar was calibrated before the experiment by using a pink noise signal that was played through a Mackie MR8 Active Studio Monitor speaker and calibrated using a Quest Technologies 1200 sound level meter (A weighting). The resulting calibration coefficient converted the rms of the windowed microphone signal to decibels of SPL, which was then used to update the display of the sound bar.

Each trial lasted 5.8 s and began with presentation of the microphone feedback signal to the insert earphones and initiation of audio recording for that trial. The word appeared 500 ms after the feedback was presented and was displayed for 1.3 s. Speakers were instructed to pronounce each word clearly and to prolong the vowel until the word disappeared from screen. One and a half seconds after the word disappeared, the auditory feedback terminated and the audio recording for that trial was saved to disk for analysis. Two and a half seconds later, the next trial began. Before the start of the experiment, speakers completed two “practice” blocks of the training and test words and received either auditory feedback or speech-shaped noise as they would during the experiment. The practice blocks provided speakers with the opportunity to practice prolonging the vowels in their word productions and using the sound bar to regulate their speech intensity. Two speakers required an additional practice block to accommodate to the speech task.

Experimental design.

Words were presented in blocks of 11 trials consisting of the 6 perturbed words and the five test words. Each speaker produced a total of 550 words over a total of 50 blocks. The order of the perturbed words and the test words was randomized in each block, but perturbed words were always presented before test words. The auditory perturbation consisted of a 130-mel increase in speakers’ vowel F1. A perturbation magnitude of 130 mels was selected for this study because it is comparable to the linear F1 perturbation used by Tourville et al. (2008), which elicited robust compensation during/ɛ/, and because Reilly and Dougherty (2013) found that F1 perturbations of 130 mels elicited robust compensations during /ʌ/ (as in “bug”). The perturbation was delivered during production of the training words over the course of three experimental phases: 1) a baseline phase consisting of 12 blocks when auditory feedback was not perturbed; 2) a ramp phase that consisted of 5 blocks, during which auditory perturbations was introduced and gradually increased; and 3) a perturbation phase consisting of 33 blocks of trials when the auditory perturbation was held constant at 130 mels. As a result, vowels for which generalization was examined (/ɛ/ and either /i/ or /æ/) were each produced a total of 66 times (33 blocks × 2 words) during the perturbation phase of the experiment. Figure 2 displays the perturbation magnitude by block and phase and also depicts the order of presentation for training and test words within each block. Subjects were given the opportunity to take a short break after every other block, during which they were allowed to stretch or drink water, as needed, without speaking.

Fig. 2.

The experiment consisted of 50 blocks of trials that were organized into 3 phases: 1) a baseline phase, when auditory feedback was not perturbed; 2) a ramp phase, during which the auditory perturbation of F1 was introduced and gradually increased; and 3) a perturbation phase, when the F1 perturbation was held constant at 130 mels. Each block consisted of 11 trials. During the first 6 trials of a block, the speaker produced a random ordering of each of the 6 training words, and auditory feedback of their speech was delivered via insert earphones. On the final 5 trials in the block, the speaker produced a random ordering of the 5 test words, and during these productions the speaker’s auditory feedback was replaced with speech-shaped noise.

Because of the study’s focus on generalization, the words and experimental design differed from those of similar adaptation studies. For example, nonperturbed vowels were produced in more than one word, and the perturbation phase of the experiment was extended so that a greater number of nonperturbed vowel productions would be available for analysis. In addition, the decision to embed each perturbed vowel in a variety of phonetic contexts (i.e., 6 different words) was motivated by visuomotor findings demonstrating that perturbation to a larger portion of a workspace results in broader generalization than perturbation to smaller portion of that workspace (Neva and Henriques 2013). The different phonetic contexts were expected to increase the range of formant values produced during the transition and steady-state portions of the perturbed vowels and, in turn, expand the portion of the vowel space exposed to the perturbation. To offset the longer perturbation phase, a postperturbation phase was not included in the study. As a result, the findings of the present study are limited to vowel changes associated with adaptation and generalization. Vowel changes associated with the return to baseline from those states have the potential to inform about vowel coupling but were not addressed in the present study.

Analysis.

A custom MATLAB graphic user interface (GUI) was used to display the speech microphone signal, the Audaptor-derived F1 and F2 frequency contours, and the broadband spectrogram of the microphone signal for each trials. For trials containing F1 perturbations, the perturbed F1 signal was also displayed. Using this information, trained users identified the onset and offset of the steady-state portion of the vowel, from which the mean F1 and F2 frequencies were derived. The steady state of the vowel corresponded to the set of F1 and F2 values following the F2 transition out of the initial stop consonant and before the F2 transition into final stop consonant. The bottom of the GUI displayed a spectrogram with the Audapter-derived values overlaid. A comparison of the formant regions in the spectrogram with the Audapter-derived formants allowed for identification of overt errors in formant tracking. These errors, which amounted to ~6% of all trials, were typically due to the identification of a harmonic as a formant. All such errors were discarded from analysis.

Average F1 and F2 values for each vowel in the training and test lists were derived for each of the 50 blocks in the study. These F1 and F2 averages for each vowel were expressed relative to each speaker’s baseline formants by subtracting the mean of the F1 and F2 averages during the 12 baseline blocks from the F1 and F2 averages during the other blocks of trials. As a result, F1 and F2 block averages for each vowel had a mean of zero during the baseline phase, and the block averages during the ramp and perturbation phases denoted deviations from each speaker’s baseline. Adaptation was evaluated for each speaker by testing whether the array of F1 block averages for the training vowel was significantly less than zero (P < 0.01) during the perturbation phase of the experiment. Speakers who exhibited significant adaptation were termed “adapt” speakers to distinguish them from the other participants in the study. In addition, the possibility that some speakers “followed” the perturbation was evaluated by testing whether F1 block averages during the perturbation phase were significantly greater than zero (P < 0.01).

In the present study, the term “adaptation” was used to refer to changes in the perturbed vowel during both training and test words. This is not entirely accurate because changes in training words involve a combination of adaptive and compensatory responses (Houde and Jordan 1998). Nevertheless, this characterization of adaptation was preferred because vowel changes in training words were the basis for identifying adaptation in individual speakers. Consequently, adaptations during training and test words were treated as separate variables in all analyses.

The mean F1 and F2 for each vowel was derived from the block averages during the perturbation phase and denoted the magnitude of adaptation for training and test words as well as the magnitude of generalization for the two nonperturbed vowels. For the adapt speakers, multiple t-tests were performed to evaluate 1) whether adaptation of the perturbed vowel in training words was accompanied by changes in the perturbed vowel during test words and 2) whether adaptation generalized to nonperturbed vowels in the test words. To control for type I error, the false discovery rate (FDR) method (Benjamini and Hochberg 1995) was used to calculate a corrected significance threshold for the six t-tests in this analysis. The corrected threshold for identifying significant differences was set to P_FDR <0.05.

Differences in adaptation and generalization between adapt speakers in the pert_i and pert_æ groups were evaluated for both training and test productions of the perturbed vowel. Because generalization was not assessed for the same vowels in each speaker group, nonperturbed vowels were grouped according to their distance, either “near” or “far,” from the perturbed vowel. The near vowel corresponded to the vowel immediately adjacent to the perturbed vowel and was /ɛ/ for both speaker groups. The far vowel corresponded to the vowel that was further from the perturbed vowel (either /i/ or /æ/). Generalization differences between the pert_i and pert_æ groups were then assessed by comparing the magnitude of changes in the near and far vowels.

Correlation analyses evaluated whether adaptation varied predictably with generalization in nonperturbed vowels and whether generalization in the latter vowels varied predictably with each other. For these analyses, data were collapsed across the two speaker groups and six correlations were evaluated. The correlation between F1 adaptation during training words and test words assessed whether these variables measured related aspects of adaptation. Correlations between F1 adaptation during training and test words and generalization in the near and far nonperturbed vowels assessed whether adaptation was linearly related to generalization. Finally, the correlation between generalization in the near and far vowels assessed whether changes in one nonperturbed vowel were predictably related to changes in another nonperturbed vowel. This set of six correlations was performed over the subset of speakers who exhibited significant adaptation (adapt speakers) as well as the full set of speakers who participated in the study. These correlations analyses were designed to evaluate 1) whether production of neighboring vowels covaried over a range of positive and negative F1 changes and 2) whether covariation among neighboring vowels could account for vowel generalization. Because this analysis involved 12 tests of correlation, adjusted P values were calculated using the FDR method, and a threshold of P_FDR < 0.05 was used to determine statistical significance.

RESULTS

Adaptation and generalization in each speaker group.

In the pert_i group, speakers were exposed to repeated perturbation of F1 during production of words containing the vowel /i/. Adaption was evaluated for each speaker by testing for significant F1 decreases during the perturbation phase compared with the baseline phase. This analysis identified significant adaptation (P < 0.01) in 12 of the 20 speakers in the pert_i group and in 12 of the 19 speakers in the pert_æ group. The average magnitude of adaptation was 39 mels for the speakers in the pert_i group and 38 mels for the speakers in the pert_æ group. The top row of Fig. 3A displays the F1 group means and SE by block for the perturbed vowel /i/ (left) and the nonperturbed vowels /ɛ/ (middle) and /æ/ (right). Vertical lines separate the baseline, ramp, and perturbation phases of the experiment. Figure 3B (top) displays the means and SE for each block of trials for the vowel /i/ in the test word “pick.” The bottom row of Fig. 3A displays the effects of F1 perturbations on productions of the vowel /æ/ in the pert_æ group. The mean and SE of this group’s F1 responses are shown by block for the perturbed vowel /æ/ (right) and the nonperturbed vowels /ɛ/ (middle) and /i/ (left). Figure 3B (bottom) displays the corresponding results for /æ/ during production of the test word “pack.” Changes in F2 were not detected in the adaptation responses for either speaker group (P > 0.15) and are not displayed.

Fig. 3.

Adaptation and generalization during F1 perturbations of the vowels/i/ and /æ/ in the pert_i and pert_æ groups. A: F1 means and SE by block are displayed for the pert_i group (top) and pert_æ group (bottom) during each block of trials for the vowels /i/ (left), /ɛ/ (middle), and /æ/ (right). The baseline, ramp, and perturbation phases of the experiment are denoted by dashed vertical lines in each panel. F1 averages during the baseline phase were subtracted from the values in each panel to highlight changes in these formants over the experiment. In the pert_i group (top), adaptation during training words is indicated by changes in the production of the perturbed vowel, /i/, and generalization is indicated by changes in the production of the nonperturbed vowels, /ɛ/ and /æ/. In the pert_ æ group (bottom), adaptation during training words is indicated by changes in the production of the perturbed vowel, /æ/, and generalization is indicated by changes in the production of the nonperturbed vowels, /ɛ/ and /i/. B: F1 changes in the productions of the vowel /i/ in the test word “pick” (top) and in the vowel /æ/ during the test word “pack.” Changes in the production of these vowels were used to derive speakers’ adaptation during test words.

Figure 4 displays the means and SE of vowel changes in perturbed and nonperturbed vowels among adapt speakers in each group. F1 changes from the pert_i group are displayed in black, and F1 changes from the pert_æ group are displayed in gray. Changes in the production of the perturbed vowel during test words were examined to assess whether vowel adaptation during training words was retained when auditory feedback was masked by noise. In the pert_i group, this analysis revealed significant decreases in F1 during production of the vowel /i/ in the test word “pick” [t(11) = −5.52, P_FDR < 0.001], and the average magnitude of these F1 decreases was 25 mels. In the pert_æ group, decreases in F1 were observed in the vowel /æ/ during productions of the test word “pack” [t(11) = −5.95, P_FDR < 0.001], and their average magnitude was 26 mels. These findings indicate that adaptation of the perturbed vowel was retained when auditory feedback was masked by noise. On average, retention during test words was 78% of that observed during training words, which is comparable to the 70% retention reported by Houde and Jordan (1998) but somewhat larger than the 58% reported by Villacorta et al. (2007).

Fig. 4.

Means and SE of vowel changes by group for each of the 4 vowels evaluated in this study. Results from the pert_i group are displayed in black, and results from the pert_æ group are displayed in gray.

An analysis of F1 changes in nonperturbed vowels was performed to determine whether speakers’ adaptation to the perturbation generalized to other, nonperturbed vowels. In the pert_i group, adaptation to the perturbation of the vowel /i/ was accompanied by similar changes in the production of /ɛ/ and /æ/. That is, significant decreases in F1 were observed during production of words containing /ɛ/ [t(11) = −3.43, P_FDR < 0.01] and /æ/ [t(11) = −2.91, P_FDR < 0.05] during the perturbation phase of the experiment. The average decrease in F1 was 18 mels for /ɛ/ and 17 mels for the /æ/.

In the pert_æ group, speakers who adapted to the perturbation of words containing /æ/ exhibited significant generalization in the nonperturbed vowels, /ɛ/ and /i/. F1 decreased during /ɛ/ [t(11) = −5.60, P_FDR < 0.001] by an average of 25 mels and during /i/ [t(11) = −3.43, P_FDR < 0.01] by an average of 15 mels.

In summary, these findings regarding non-perturbed vowel changes indicate that adaptation to perturbation of either /i/ or /æ/ generalized to nearby vowels that were not perturbed.

Between-group comparisons of adaptation and generalization.

A repeated-measures ANOVA was performed to test for group differences in the magnitude of adaptation observed in the training and test words. For this analysis, word type (training vs. test) was a within-subjects factor and group (pert_i vs. pert_æ) was a between-subjects factor. This analysis did not identify a main effect of group [F(1, 22) = 0.001, P = 0.98] or an interaction between group and word type [F(1, 22) = 0.085, P = 0.77]. A main effect of word type was identified [F(1, 22) = 9.21, P < 0.01] and indicated that perturbed vowel changes were greater during the training words than during the test words by an average of 13 mels.

The generalization magnitudes of the two speakers groups were also compared using a repeated-measures ANOVA. For this test, distance (near vs. far) was a within-subjects factor and group (pert_i vs. pert_æ) was a between-subjects factor. This analysis did not identify a difference in the magnitude of generalization between the two speaker groups [F(1, 22) = 0.198, P = 0.66]. In addition, no effect of distance [F(1, 22) = 2.39, P = 0.14] and no interaction between group and distance [F(1, 22) = 1.54, P = 0.23] were present.

Following responses.

In addition to the 24 speakers who adapted to the perturbation, 2 speakers in the pert_i group and 3 speakers in the pert_æ group demonstrated following responses to the perturbation (P < 0.01). The average magnitude of the following response was 27 mels in the pert_i group and 26 mels in the pert_æ group, and no between-group difference was present in the magnitude of following responses [t(3) = −1.78, P = 0.17]. Neither group exhibited significant changes in F2 during the perturbation phase (P > 0.27).

Despite the small sample sizes, a preliminary analysis was conducted to evaluate whether the following responses were accompanied by vowel changes in the test words. Data were combined from the two speaker groups to partly offset the small sample sizes for this analysis. Given the exploratory nature of these analyses, P values were not corrected for the use of multiple t-tests. This analysis identified significant increases in F1 during production of the perturbed vowel in test words [t(4) = 2.96, P < 0.05]. The average magnitude of these F1 increases was 15 mels. This analysis did not identify significant generalization to either the near vowel [t(4) = 1.14, P = 0.32] or the far vowel [t(4) = 0.40, P = 0.71].

Correlation testing.

The absence of between-group differences in adaptation precluded examination of connections between adaptation and generalization at the group level. Instead, the question of whether or how adaptation differences affected generalization was evaluated using correlation tests of within-speaker F1 changes across vowel pairs. Figure 5 displays scatterplots corresponding to each test of correlation evaluated below. The top row of Fig. 5 depicts the relation between perturbed vowel changes during training words and near vowel changes (left), far vowel changes (middle), and perturbed vowel changes during test words (right). The bottom row of Fig. 5 depicts the relation between perturbed vowel during test words and near vowel changes (left) and far vowel changes (middle), and the relation between near and far vowel changes is depicted (right) In each panel, the data from all speakers are depicted by the open black circles and the subset of data from adapt speakers are indicated with gray circles. For the adapt speakers’ data, the perturbed vowel changes during training and test words denote adaptation and the near and far vowel changes denote the corresponding generalization. The lines of best fit for all speakers’ data are shown in black, and the lines of best fit for the adapt speakers are shown in gray.

Fig. 5.

Scatterplot displays corresponding to each of the correlation tests evaluated in the study. Each panel displays the data for all speakers in the study (open black circles) as well as the subset of data from speakers who adapted to the auditory perturbation (filled gray circles). The lines of best fit and squared correlation values for the comparisons involving all speakers’ data are shown in black, and those involving only speakers who adapted are shown in gray.

In the adapt set of speakers, adaptation during training words was not significantly correlated with adaptation during test words (r = 0.27, P_FDR = 0.25) or with generalization in the near vowel (r = 0.18, P_FDR = 0.42) or the far vowel (r = −0.09, P_FDR = 0.68). Adaptation during test words exhibited significant correlations with generalization in the near vowel (r = 0.47, P_FDR < 0.05) but not the far vowel (r = 0.35, P_FDR = 0.12). A significant correlation (r = 0.52, P_FDR < 0.05) was observed between the near and far vowel generalization.

In the analysis of data from all speakers, F1 changes in the perturbed vowel during training words were significantly correlated with F1 changes during test words (r = 0.74, P_FDR = 0.0001) and were also correlated with F1 changes in the near vowel (r = 0.56, P_FDR < 0.001) but not the far vowel (r = 0.33, P_FDR = 0.07). In addition, F1 changes in the perturbed vowel during test words were significantly correlated with F1 changes in the near vowel (r = 0.68, P_FDR < 0.0001) and the far vowel (r = 0.43, P_FDR < 0.05). Changes in the near and far vowels were also significantly correlated (r = 0.58, P_FDR < 0.001).

DISCUSSION

The present study investigated the mechanisms underlying vowel generalization and their relation to sensorimotor representations of vowels. Perturbation of /i/ and /æ/ did not elicit differences in either adaptation or generalization among speakers who exhibited significant adaptation to the perturbation. Analyses of vowel changes between the baseline and perturbation phases revealed significant correlations between the F1 changes of adjacent vowels but not nonadjacent vowels. The findings indicate that vowel changes associated with generalization are due in part to control parameters that link adjacent vowel representations, possibly to maintain their acoustic contrast, or separation.

Adaptation to repeated perturbations of vowel auditory feedback has been well documented and highlights the interactive nature of feedforward and feedback control in contemporary models of speech production (Guenther and Perkell 2004; Hickok et al. 2011; Houde and Nagarajan 2011; Tourville et al. 2014; Villacorta et al. 2007). The present study found that 24 of 39 speakers adapted to (i.e., opposed) the perturbation and that 5 speakers followed the perturbation. Although opposing, or compensatory, responses to an auditory perturbation are more common, reports of following responses to perturbations of F0 (Burnett et al. 1998; Munhall et al. 2009) and vowel formants (Cai et al. 2010; MacDonald and Munhall 2012; Munhall et al. 2009; Villacorta et al. 2007) are sufficient in number to indicate that they are not outliers. Moreover, speakers in this study who exhibited following responses during training words also exhibited following responses when producing the perturbed vowel in test words. This finding indicates that following responses, like adaptive responses, are learned. Although following responses are partly attributable to the properties of the perturbation (Burnett et al. 1998), these responses may also reflect individual differences in feedback control for speech (Larson and Robin 2016; MacDonald and Munhall 2012). Further research is needed to elaborate the particular differences in feedback control and whether they are related to individual differences in speakers’ vowel spaces.

Among the speakers who adapted to the perturbation, the average change in F1 was 38 mels in the pert_i group and 39 mels in the pert_æ group. No difference in adaptation magnitude was detected between the two speaker groups. This finding was unexpected given previous findings demonstrating the sensitivity of compensation and adaptation responses to the locations of neighboring vowels (Mitsuya et al. 2011, 2013; Reilly and Dougherty 2013). In the present study, it was expected that an upward F1 perturbation would elicit larger adaptation during /i/ than during /æ/ because the former has an adjacent vowel, /ɛ/, in the direction of the perturbation and the latter does not.

One possible reason for the discrepancy between the current and previous findings is that the perturbation of F1 shifted lax productions of /æ/ to one of this vowel’s tense variants. Because allophonic variants of /æ/ are most often bound to phonetic context (Gordon 2005), the mismatch between allophone and phonetic context may have elicited a larger-than-expected corrective response. Alternatively, it may be that the contrast between /i/ and /ɛ/ may have been insufficiently small to elicit the larger adaptation associated with a neighboring vowel. For example, the F1 separation between /i/ and /ɛ/ is relatively large among adjacent vowels in the English vowel space, and, in addition, /i/ and /ɛ/ have relatively large differences in F2, which further increases their contrast distance (Hillenbrand et al. 1995; Peterson and Barney 1952).

For adapt speakers in both groups, changes to the perturbed vowel were accompanied by similar changes to nearby vowels that were not perturbed. In the pert_i group, F1 decreased by 18 mels and 17 mels for the vowels/ɛ/ and /æ/, respectively, and in the pert_æ group, the vowels /ɛ/ and /i/ exhibited F1 decreases of 25 mels and 15 mels, respectively. The finding of changes in nonperturbed vowels in two different groups of speakers, each of whom received a perturbation to a different vowel, is consistent with previous findings (Cai et al. 2010; Houde and Jordan 1998; Rochet-Capellan and Ostry 2011; Villacorta et al. 2007) and provides further support for the idea that adaptive changes in perturbed vowels generalize to other, nonperturbed vowels. Generalization is also a common finding in limb motor studies (Ahmed et al. 2008; Ghahramani et al. 1996; Krakauer et al. 2000; Vetter et al. 1999) and indicates that this aspect of motor learning is shared across movement systems.

Correlation findings.

Vowel generalization and other findings from auditory perturbation studies have led investigators to propose that feedforward representations (goals) for vowels are specified jointly, rather than independently, such that the target regions for any particular vowel depend on the target regions of other vowels (Cai et al. 2010; Houde and Jordan 1998; Mitsuya et al. 2013). If generalization results from vowel representations that are at least partly shared, then one would expect changes associated with adaptation to be correlated with changes associated with generalization. In addition, one would also expect that generalization among different nonperturbed vowels to be correlated.

The correlation findings in the adapt speakers produced discrepant findings regarding this question. For example, adaptation during training words was not correlated with either near or far vowel generalization. However, adaptation during test words was correlated with near vowel generalization, and generalizations of the near and far vowels were also correlated.

The discrepancies in the correlation findings of the adapt speakers is likely related to the different speaking conditions in this study. Adaptation during training words was evaluated when auditory feedback was present, but adaptation and generalization during test words were evaluated in the presence of masking noise that minimized auditory feedback. The different speaking conditions during training and test productions likely influenced correlation results in ways that were not related to the mechanisms underlying adaptation and generalization. For example, adaptation during training words is supplemented by ongoing feedback-based compensations to the perturbation (Houde and Jordan 2002). At the same time, adaptation and generalization during test words could have been subject to noise-induced increases in formant values (Bond et al. 1989; Garnier and Henrich 2014; Perkell et al. 2007; Van Summers et al. 1988). This could have happened if the virtual sound bar was not wholly effective in controlling for speech intensity differences or if the noise elicited changes in F1 that were independent of changes in intensity. The strongest evidence that aspects of the different speaking conditions influenced the correlations findings is the absence of a correlation between changes in the perturbed vowel during training and test productions. Additional evidence of a speaking condition effect is indicated by the finding that the three correlation tests examining a relationship between test and training words produced the three weakest correlations in the study.

Correlations involving the perturbed vowel in test words were not subject to either speaking condition effects and can therefore offer insight into the relationship between changes in perturbed and nonperturbed vowels. These findings revealed that adapt speakers exhibited a significant positive correlation between adaptation during test words and near vowel generalization but not far vowel generalization. In addition, changes in the near vowel were positively correlated with changes in the far vowel. Together, correlations among the test words indicate that vowel changes were correlated between adjacent vowels but not nonadjacent vowels.

The results of the correlation testing in all speakers were consistently higher than those of the adapt speakers. This likely reflects the increased range of values over which the correlations were evaluated (Glass and Hopkins 1970; Goodwin and Leech 2006). Nevertheless, the findings of the correlation tests from all speakers’ data were consistent with several of the correlation findings from the adapt speakers. For example, both analyses identified significant correlations between changes in the perturbed vowel during test words and changes in the near vowel, /ɛ/. In addition, near vowel changes were significantly correlated with far vowel changes in both analyses. That these correlations were significant in both analyses indicates that coupling between adjacent vowels pairs was not restricted to the F1 decreases associated with adaptation. Rather, this coupling appeared to be a more general feature of vowel production that linked adjacent vowels over a range of changes in F1. At the same time, the findings that correlations between perturbed vowel changes and far vowel changes were rarely observed indicates that this coupling does not extend to nonadjacent vowels.

The present results indicate that coupling between the representations of adjacent vowels is partly responsible for vowel generalization. As a result, these findings indicate that generalization is not entirely a consequence of the auditory perturbation but is partly due to the adaptation to that perturbation. The notion that factors other than the sensory perturbation contribute to generalization is consistent with findings from the limb motor literature. For example, in their study of visual perturbations during reaching, Taylor and Ivry (2013) reported different patterns of generalization that depended on the arrangement of test targets in the workspace. Reichenthal et al. (2016) found that goal requirements, which corresponded to the size of targets during training, also elicited differences in generalization. Of particular relevance to the current study are the findings by Alexander et al. (2013), who evaluated participants’ adaptation to prism glasses during different walking tasks. These investigators found significant differences in generalization that depended on whether or not generalizing resulted in the avoidance of a negative consequence (i.e., collision with an obstacle).

The findings of Alexander et al. (2013) indicate that behavioral changes associated with generalization are modulated by their functional consequences. In the case of vowel production, the functional consequences of generalization may relate to the separation, or contrast, between the perturbed vowel and neighboring vowels. Absent generalization, adaptation will typically cause the perturbed vowel to sound similar to a neighboring vowel, which would adversely impact speech intelligibility (Bradlow et al. 1996; Kim et al. 2011; Lansford and Liss 2014; McCaffrey and Sussman 1994; Moon and Lindblom 1994; Nakamura et al. 2008; Picheny et al. 1986; Turner et al. 1995). The notion that separability, or contrast, between neighboring vowels can influence speech motor output is also consistent with speech production findings (Manuel 1990; Perkell et al. 2007).

At the same time, the present findings also indicate that coupling between adjacent vowels is not sufficient to account for generalization. Although the correlations among vowels in the test words were significant, the magnitude of these correlations indicates that changes in one vowel still accounted for less than half of the variance of changes in an adjacent vowel. As a result, it seems likely that other mechanisms, more directly attributable to the perturbation itself, also contributed to generalization. Several contemporary models (Donchin et al. 2003; Ghahramani et al. 1996) link generalization directly to the sensory perturbation via error-driven processes that effectively generate a local remapping of sensory-to-motor space in the region of the perturbation. These models account for a range of findings regarding generalization, including some findings regarding vowel generalization (Rochet-Capellan et al. 2012). Further research is necessary to elaborate the interaction between error-driven processes and inter-vowel dependencies as they relate to vowel generalization.

GRANTS

This research was supported by National Institute on Deafness and Other Communication Disorders Grant R03 DC011159.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

K.J.R. conceived and designed research; K.J.R. and C.P. performed experiments; K.J.R. and C.P. analyzed data; K.J.R. and C.P. interpreted results of experiments; K.J.R. prepared figures; K.J.R. drafted manuscript; K.J.R. and C.P. edited and revised manuscript; K.J.R. and C.P. approved final version of manuscript.