Age-related changes to spectral voice characteristics affect judgments of prosodic, segmental, and talker attributes for child and adult speech

Abstract

Purpose

As children mature, changes in voice spectral characteristics covary with changes in speech, language, and behavior. Spectral characteristics were manipulated to alter the perceived ages of talkers’ voices while leaving critical acoustic-prosodic correlates intact, to determine whether perceived age differences were associated with differences in judgments of prosodic, segmental, and talker attributes.

Method

Speech was modified by lowering formants and fundamental frequency, for 5-year-old children’s utterances, or raising them, for adult caregivers’ utterances. Next, participants differing in awareness of the manipulation (Exp. 1a) or amount of speech-language training (Exp. 1b) made judgments of prosodic, segmental, and talker attributes. Exp. 2 investigated the effects of spectral modification on intelligibility. Finally, in Exp. 3 trained analysts used formal prosody coding to assess prosodic characteristics of spectrally-modified and unmodified speech.

Results

Differences in perceived age were associated with differences in ratings of speech rate, fluency, intelligibility, likeability, anxiety, cognitive impairment, and speech-language disorder/delay; effects of training and awareness of the manipulation on ratings were limited. There were no significant effects of the manipulation on intelligibility or formally coded prosody judgments.

Conclusions

Age-related voice characteristics can greatly affect judgments of speech and talker characteristics, raising cautionary notes for developmental research and clinical work.

Prosody is an important aspect of verbal communication that children must master to achieve adult-like speech-language competency. Prosodic cues such as pitch and timing are important for conveying a variety of information in spoken language, including semantic, lexical, syntactic, and emotional information (Lehiste, 1970; Scherer, 2003). Children’s ability to effectively use prosody to communicate improves over time (Allen & Hawkins, 1980; Snow, 1995). Assessment of children’s and adults’ verbal communicative competency frequently involves characterizing prosodic attributes (e.g., speech rate, pausing, phrasing, and fluency) alongside speaker attributes (e.g., developmental age, language skill, and personality factors) in order to identify errors and develop treatment goals.

This paper investigates perception of prosodic, segmental, and talker attributes, focusing on whether judgments about such attributes are dependent on a listener’s knowledge or beliefs about a talker – particularly the talker’s probable age – and whether such dependence varies by assessment method. In addition to having clinical and research applicability, this investigation addresses the theoretical issue of whether prosodic (i.e., suprasegmental) judgments of speech are independent of other components of the speech signal, as predicted by theories of speech processing in which suprasegmental, segmental, and indexical components of the signal are treated as modular (Halle, 1985; Kuhl, 1991).

Accurate characterization of prosody is important for understanding speech-language development in children, as well as for assessing speech-language disorders in both children and adults. Prosody is often impacted in a variety of communication disorders (Peppe, 2009), and speech-language pathologists (SLPs) must judge prosodic attributes such as speech rate, pausing/phrasing, and fundamental frequency (F0) as part of diagnosing and treating speech-language disorders. Both in clinical and research settings, prosody judgments are often made impressionistically using ad-hoc approaches, as opposed to using more formalized coding systems, rating procedures, or acoustic measurements.

Prior research suggests impressionistic judgments of prosody (e.g., speech rate, F0) may be susceptible to listener biases. For example, Bond and colleagues (Bond, Simpson, & Feldstein, 1988) showed that listeners judged the same acoustic speech rate as faster or slower when presented at different levels of F0 or intensity. This suggests that perceived speech rate is not merely a function of acoustic rate but depends on complex, inter-related acoustic factors, a proposition bolstered by similar findings in nonspeech auditory perception (Henry, McAuley, & Zaleha, 2009; Melara & Marks, 1990). If judgments of one attribute (e.g., speech rate) depend not only on the acoustics of that attribute (e.g., acoustic speech rate), but also another (e.g., F0), then therapy might inadvertently target suboptimal characteristics. For example, if a talker’s speech rate is perceived as too fast or too slow, then it may be more appropriate under some conditions to target F0, rather than articulation rate per se, due to the interdependence of these acoustic attributes in perception (Bond et al., 1988; Henry et al., 2009; Melara & Marks, 1990).

Given that perceptual judgments of prosody may be subject to interdependencies of acoustic factors, a variable worthy of study is developmental age. Little research has examined the possibility of biases affecting assessment of children’s speech. However, the presumed age and gender of a child have both been shown to influence perceptions of children’s speech and ratings of accuracy and quality of its production (Munson, Edwards, Schellinger, Beckman, & Meyer, 2010; Munson & Seppanen, 2009).

Understanding how prosodic competency changes over time is a long-term goal of our research. Given prior findings of perceptual interdependence among acoustic variables, impressionistic measures of prosody may rely on confounded acoustic variables that change as a child matures. If so, then changes attributed to prosodic development might reflect development in other areas. Thus, it is important to know if perception of prosodic attributes (e.g., perceived speech rate) is dependent merely on simple acoustic properties (e.g., acoustic speech rate) or a broader set of quasi-orthogonal acoustic characteristics.

Many acoustic characteristics change concurrently as children age, further underscoring the possibility of confounds between perception of prosodic and other acoustic attributes. For example, speech rate increases from ages 5–15 (Lee, Potamianos, & Narayanan, 1999). F0 tends to decrease with age, even in prepubescent children (Sussman & Sapienza, 1994). Likewise, formants tend to lower as children develop (Lee et al., 1999). Given findings that acoustic attributes underlying prosodic cues are perceptually interdependent (Bond et al., 1988), it is important to determine how perception of prosody may be affected by systematic changes in acoustic variables correlated with developmental age. Indeed, the converse relationship holds: for adults’ speech, judgments of talker age are affected by manipulations to F0, F1, and/or speech rate (Harnsberger, Shrivastav, Brown, Rothman, & Hollien, 2008; Reubold, Harrington, & Kleber, 2010). However, no prior published work has examined the effect of F0 or formant manipulations on judgments of talker age for children’s speech.

The critical manipulation in our experiments was a spectral manipulation of speech harmonics and formant frequencies which was expected to affect perceived talker age and possibly prosody judgments, if such are confounded with other perceptual characteristics of a speaker’s voice. This spectral manipulation, which was expected to make children sound older, and adults younger, critically permitted us to examine whether judgments of prosody would be affected by age-related voice characteristics, while holding constant key acoustic-prosodic variables: acoustic speech rate; frequency, duration, and number of pauses; and the pattern of F0 ups and downs (i.e., the F0 contour), which conveys a variety of linguistic information (Lehiste, 1970). The comparison of greatest interest therefore involved judgments of prosody for a talker’s speech under different spectral manipulation conditions where the talker was perceived as older or younger. Exp. 1a compared such judgments for naïve participants and for participants who were told that the speech had been altered so talkers may sound older or younger. Likewise, Exp. 1b examined judgments made by SLP Master’s students of speech, to determine if individuals with relatively more training in speech-language pathology would also be susceptible to age-related bias in prosody judgments.

By comparing judgments of different listeners, questions could be addressed regarding how awareness of potential bias and/or learning and experience might each affect judgments. The first question addressed was whether age-related bias in prosody judgments is relatively automatic, or alternatively, whether age-related bias is a function of awareness. Prior research suggests that even when individuals are aware of biases (e.g., in racial attitudes), there can be long-lasting, subtle effects on processing (Dasgupta & Greenwald, 2001). Exp. 1a therefore compared judgments between listeners who were made aware of the spectral age manipulation with listeners who were not informed of this manipulation. The second question addressed was whether individuals with greater training and experience in speech development and assessment – here, SLP Master’s students – are susceptible to age-related bias in judgments of prosody and other characteristics. Less age-related bias might be expected for the SLP Master’s students than for the psychology student undergraduates, due to the former having substantially more experience in speech-language issues, which might permit more reliable assessment of prosody under different spectral profiles. On the other hand, speech perception is affected by socio-developmental characteristics of talkers, such as age and social class (Drager, 2011; Hay, Warren, & Drager, 2006; Johnson, Strand, & D’Imperio, 1999; Munson, Jefferson, & McDonald, 2006), suggesting that greater experience with different talker groups might affect speech perception. On this basis we might expect more susceptibility to age-related bias in prosody judgments by SLP Master’s students than by undergraduate students in Exp. 1a.

Another question addressed here was how a talker’s baseline degree of speech-language competency affects prosodic and other judgments. Talkers in our studies were either 5-year-old children or adults. Five-year-olds differ from adults in many speech-language characteristics, including phonology, syntax, and vocabulary (Bernthal, Bankson, & Flipsen, 2008). We predicted that baseline speech-language skill would affect prosody judgments, such that effects of our spectral “age” manipulation on such judgments would differently affect adults’ vs. children’s speech. For example, we expected that in the unmodified condition, children’s speech rate should be judged as slower than that of adults, since acoustic rate is slower for children than adults (e.g., Lee et al., 1999). However, age-related bias about speech rate should trend in the direction of the expected acoustic speech rate, given the perceived age of the talker. Thus, children’s speech should be judged as slower in the spectrally modified (i.e., older) condition than the unmodified condition, since a child’s speech rate is age-appropriate for a child’s voice but too slow when paired with an adult’s voice; the reverse should be true for adults’ speech.

The talkers selected here were 5-year-old children and adults, due to these groups’ substantial differences in fluency and articulation (Redford, 2012; Bernthal et al., 2008). Fluency is strongly related to prosody, including duration, rate, pausing, and pitch, but is not independent of segmental pronunciation (Bloodstein, 1995). Articulation is chiefly a segmental characteristic, and greatly influences intelligibility. An advantage of our spectral manipulation method was its good preservation of segmental information, e.g., distinctive feature contrasts for consonants and vowels. Thus, our manipulation permitted investigation of age-related bias in perception of the (quasi-prosodic) attribute of fluency, as well as the (largely segmental) attribute of intelligibility. There is a dearth of research on potential age-related bias in impressionistic judgments of fluency and intelligibility. Such judgments are widely used in clinical practice (Bernthal et al., 2008; Bloodstein, 1995), even though they are notoriously unreliable and have been shown to have interdependencies with other perceptual variables (Bernthal et al., 2008; Southwood & Weismer, 1993). No research that we know of has investigated whether differences in judgments of intelligibility and fluency might be associated with differences in the perceived age of a talker.

It was also predicted that bias about age-related speech-language performance expectations would be observed in judgments of typicality or impairment in talkers. For example, a relatively high disfluency or misarticulation rate would likely be interpreted as normal for a typically developing, non-stuttering, five-year-old child. However, if the same (child) talker’s voice is made to sound like that of an adult, then an identical disfluency and misarticulation rate might be deemed as indicating speech-language impairment (e.g., stuttering or a motor speech disorder), or perhaps cognitive impairment. A complementary spectral transformation for adults’ speech was not predicted to have the same effect; since typical adults produce few disfluencies and misarticulations and have relatively high vocabulary and language skills, transforming the spectrum of (adult) speech to sound like that of a child should result in talker’s sounding extraordinarily competent (and unimpaired in speech, language, or cognition). Such a pattern of results would be consistent with a hypothesis that ratings of prosody and other speech attributes depend not only on isolated characteristics of a talker’s speech, but on the listener’s overall assessment of the talker’s characteristics of an individual, including those related to probable impairment.

Finally, we assessed two additional properties related to communication impairment. The first was the degree of perceived anxiety of the talker, based on the observation that people who stutter tend to be judged as sounding more anxious than people who don’t (Menzies, Onslow, & Packman, 1999). Child talkers in the spectrally-modified condition were predicted to be judged as sounding particularly anxious, due to the likelihood of their being judged as having a fluency-related disorder. Second, we investigated the likeability of a talker, based on findings showing that individuals with cognitive or communication impairments tend to be viewed more negatively than those without them (Franck, Jackson, Pimental, & Greenwood, 2003; Lass, Ruscello, Harkins, & Blankenship, 1993).

In summary, this research used a spectral manipulation to alter a talker’s perceived age and to determine its subsequent effects on judgments of prosody and other speech and speaker characteristics. Exp. 1 examined judgments of prosodic and articulatory properties (speech rate, fluency and intelligibility) and speaker characteristics (cognitive and/or communication impairment, anxiety, and likeability) as a function of spectral manipulation conditions by three groups of listeners: naïve undergraduates vs. undergraduates who were explicitly told about the voice alteration (Exp. 1a), and SLP Master’s students (Exp. 1b). Select comparisons among the groups permitted examination of the extent to which such judgments are susceptible to age-related bias as a function of awareness of unreliability of voice age characteristics and/or listener experience. Exp. 2 was a control study aimed at determining the extent to which intelligibility was degraded due to the spectral manipulation, to determine the extent to which age-related bias affected impressionistic intelligibility judgments in Exp. 1. Finally, Exp. 3 involved use of formal prosody labeling to determine its comparative freedom from age-related bias relative to impressionistic judgments. It was predicted that the more “analytical” method of formal prosody coding would result in smaller effects of perceived age on perception of prosody than impressionistic judgments of prosody from Exps. 1a and 1b.

Experiment 1a

Exp. 1a considered the issue of age-related bias in judgments of prosody, speech, and language and/or cognitive impairment. To investigate whether age-related bias is automatic, we compared judgments of undergraduate participants who were unaware that the talker’s speech may have been spectrally modified to a group of undergraduate participants who were made aware of the potential unreliability of voice age characteristics. If age-related bias is a function of awareness of the unreliability of voice age characteristics, then a reduction in the effects of the spectral manipulation on judgments of prosody, speech, and impairment should be observed for the group that is made aware of voice age manipulations compared with the group that is not made aware.

Methods

Participants and Design

Fifty-six undergraduate students from Michigan State University participated in the experiment in return for partial credit in a psychology course or monetary compensation. Participants were native speakers of American English who were at least 18 years of age with self-reported normal hearing. The design of the experiment was a 2 (Awareness: unaware, aware) × 2 (Talker Age: child, adult) × 2 (Modification: unmodified, modified) mixed factorial. Awareness was a between-subject factor, while Talker Age and Modification were within-subject factors. The unaware group consisted of 33 undergraduates (6 M, 27 F) with a mean age of 19.5 years, while the aware group consisted of 23 undergraduates (9 M, 14 F) with a mean age of 21.5 years.

Stimuli

Stimuli for all experiments consisted of recordings of spontaneous speech produced during the telling of one of four Frog Stories picture books by Mercer Mayer (Mayer, 1967, 1969, 1973, 1975). Speech materials were produced by nine children (3 F, 6 M; mean age: 5.5 years) and their mothers (mean age: 31 years)¹; see Redford (2012) for details. Talkers were instructed not to interrupt the person telling the story, although minor interruptions occurred. The children were all in kindergarten and were tested within the first four months of starting school. All of the children were typically developing with normal speech and hearing abilities. Recordings took place in a child-friendly, quiet experimental room in the Speech and Language Laboratory at the University of Oregon. The mean duration of the entire recording (including pauses) produced by adults was 154.6 sec and by children was 136.1 sec.

Each full recording was first segmented into 3–10 fragments (M = 5.8, total fragments = 105). Fragments for adult and child speech (including pauses) had a mean length of 25.3 sec and 24.5 sec, respectively, and contained an average of 57.6 words and 33.6 words, respectively. To make talkers for each speech file sound older or younger, we used spectral modification tools in Praat (Boersma & Weenink, 2002). The first step was to change global spectral parameters for a talker using a pitch-synchronous overlap-add (PSOLA) algorithm (Moulines & Charpentier, 1990). This involved setting a global value for each talker for two pitch measurement factors (i.e., the pitch floor and pitch ceiling) and two modification parameters (i.e., formant shift ratio and new pitch median); two other modification parameters (i.e., pitch range factor and duration factor) were left at the default value of 1.0 for all talkers. Specific values of global settings for the pitch measurement factors and modification parameters were restricted to particular ranges for adult vs. child speech. Values were selected for which the talker seemed to be most convincingly transformed in age such that children’s voices sounded maximally adult-like, and adults’ voices sounded maximally child-like and for which the speech was relatively free of artifacts. For the adults’ speech, the F0 and formants were raised by a separate fixed factor for each speaker to values judged to make the speech sound as child-like as possible; in particular, the formant shift ratio was set to a value greater than 1.0, and adjustments were made to the new pitch median, the pitch floor, and the pitch ceiling (see Table 1 for details). For children’s speech, the F0 and formants were lowered by a separate fixed factor for each speaker to values judged to make the speech of each talker sound adult-like; in particular, the formant shift ratio was set to a value less than 1.0, and adjustments were made to the new pitch median, the pitch floor, and the pitch ceiling (see Table 1). To create the most natural-sounding voices, the speech of all eight children was transformed into an adult male voice. The modifications left intact key prosodic attributes, including speech rate, F0 contour, pause timing and length across spectral modification conditions.

Table 1.

Spectral modification parameters for experimental stimuli.

	Formant Shift Ratio		New Pitch Median (Hz)		Pitch Floor (Hz)		Pitch Ceiling (Hz)
Nature of Modification	Mean (SD)	Range	Mean (SD)	Range	Mean (SD)	Range	Mean (SD)	Range
Adult → Child	1.2 (0.03)	1.2–1.3	305.0 (32.8)	265–360	96.7 (22.8)	75–150	632.2 (33.5)	600–700
Child → Adult	0.7 (0.05)	0.6–0.7	140.6 (34.8)	100–210	83.3 (14.8)	60–100	520.0 (113.7)	320–610

Following selection of global spectral parameters, the speech sounded very natural, with little if any perceptual indication of modification. To further minimize possible artifacts of modification, the speech was carefully checked auditorily. Occasional pitch errors were dealt with by hand-correction of the F0 contour using PSOLA in Praat. The few other artifacts that occurred (i.e., crackling sounds or other unnatural-sounding isolated sections) were dealt with by (i) splicing out the short portion of modified speech evidencing the artifact, (ii) subjecting the corresponding speech portion from the unmodified speech file to a different set of spectral modification parameters using a trial-and-error approach and obtaining the most natural-sounding result, then (iii) splicing the result back into the modified file.

To maintain a reasonable experiment length, we used a subset of the stimuli for Exp. 1a and 1b; all of the stimuli were used in Exp. 3. Two fragments were chosen from each child and adult’s retelling of the story which were at least 20 sec long and contained at least 25 words. When multiple fragments met these qualifications, the two which contained the most words were chosen. This yielded 36 fragments (18 talkers × 2 fragments); each fragment furthermore occurred in both unmodified and modified forms. Selected fragments were 20.5–30.0 sec long (M = 27.3) and contained 25–89 words (M = 54.5).

Four stimulus lists were created. First, the 36 speech fragments were divided into two blocks of 18 fragments; the first block contained one fragment from each of the 18 talkers, and the second block a second fragment from each talker. To create the first list, the 18 fragments within each block were arranged in a single, random order. Each fragment in the first block was then assigned pseudorandomly to either the unmodified or modified condition, while the complementary fragment by the same talker in the second block was assigned to the other condition. Sequencing and pairing were constrained so that the same Modification condition did not appear more than four times in a row and the same combination of levels of Talker Age and Modification did not occur more than three times in a row. A second stimulus list was created by inverting the Modification condition from the first list. Two more lists were created by reversing the order of the first two lists. Each fragment occurred only once on a list, so participants never heard the same fragment in both unmodified and modified forms.

Task

Participants listened to a speech fragment and estimated the talker’s age (0–100 years), and then provided judgments of seven measures while the scale for each question was displayed on the screen. (i) Speech rate was assessed on a 1–6 scale (1 = very slow, 6 = very fast). (ii) Fluency judgments were on a 1–8 scale (1 = always fluent, 8 = always disfluent). Participants were told that the term ‘fluency’ referred to the broad spectrum of fluent speech characteristics, whether by normal talkers or people who stutter, while ‘disfluency’ referred to the broad spectrum of disfluent speech characteristics.² (iii) Intelligibility was defined as how understandable the talker was; participants gave an overall impressionistic estimate from 0–100 of the percentage of words they thought they could understand. (iv) For cognitive impairment, participants categorized talkers as either probably having normal cognitive abilities or probably having a cognitive impairment. (v) For speech-language impairment, participants categorized talkers as probably having normal speech-language skills, probably having delayed speech-language skills, or probably having a speech-language disorder. For the purposes of later data analysis, the latter two categories were collapsed into a single category: ‘speech-language disorder or delay’, given the difficulty in reliably distinguishing these categories by individuals trained in communication disorders (Bernthal et al., 2008), let alone laypersons. (vi) For likeability, participants judged the likeability of the talker on a 1–6 scale (1 = extremely likeable, 6 = extremely dislikeable). (vii) Finally, participants judged the perceived anxiousness of the talker on a 1–6 scale (1 = extremely anxious, 6 = extremely calm).

Procedure

Participants in the unaware condition estimated the talker’s age and rated each of the seven attributes (see above) for each fragment. Participants in the aware condition were first familiarized with the set of speech fragments to be rated by estimating the talker’s age for each fragment. After providing age estimates for fragments, participants in the aware condition were explicitly told the speech may have been modified so that talkers sounded older or younger than they actually were. They were then told they would hear the set of speech fragments again and they should ignore perceived age while making ratings of attributes. Approximately equal numbers of participants were randomly assigned to each of the four stimulus lists in the unaware and aware conditions. Before each experiment participants in both conditions completed three practice trials using the same unmodified speech fragments so that they could become accustomed to each of the questions they would be asked during the session. Throughout the experiment, the scale associated with each question was displayed on the screen in order to reinforce consistent use of scale endpoint referents.

Apparatus

Participants listened to the speech fragments over Sennheiser HD 280 headphones. The experiment was presented to participants using E-Prime v2.0 Professional (Psychology Software Tools, Inc.) running on a Lenovo Intel® Core™2 Duo CPU E8500 with a 19-inch monitor.

Results

A 2 (Awareness: unaware, aware) × 2 (Talker Age: child, adult) × 2 (Modification: unmodified, modified) mixed-measures ANOVA was conducted for each dependent variable. The ANOVA results are summarized in Table 2. Age estimates are shown in Figure 1A–B. There were significant main effects of both Talker Age and Modification; the spectral modification was highly successful in creating different perceived ages for both Awareness conditions. Unmodified speech sounded age-appropriate (child talkers, M = 6.3 years; adult talkers, M = 29.2 years), while modifying speech made child talkers sound older (M = 19.4 years) and adult talkers sound younger (M = 13.6 years), as evidenced by the significant interaction between Talker Age and Modification. Effects of modification on estimated talker age were significant for both child talkers, t(55) = 18.0, p < .001, and adult talkers, t(55) = −28.9, p < .001. The three-way interaction between Awareness, Talker Age, and Modification reveals that having participants first estimate talker age for all fragments before making ratings of the different attributes yielded slightly more conservative age estimates (i.e., estimates that were closer to the actual ages of the talkers).

Table 2.

Summary of results of three-way ANOVAs for Exp. 1a with factors of Talker Age, Modification, and Awareness.

	Main Effect			Interactions
Dependent Variable	Talker Age	Modification	Awareness	Talker Age × Modification	Awareness × Talker Age	Awareness × Modification	Awareness × Talker Age × Modification
Age	F(1,54) = 396.80 p < .001 ηp2 = .88	F(1,54) = 9.24 p = .004 ηp2 = .15	NS	F(1,54) = 781.40 p < .001 ηp2 = .94	NS	NS	F(1,54) = 5.90 p = .019 ηp2 = .10
Speech Rate	F(1,54) = 325.40 p < .001 ηp2 = .86	F(1,54) = 18.18 p < .001 ηp2 = .25	NS	F(1,54) = 143.20 p < .001 ηp2 = .73	NS	NS	NS
Fluency	F(1,54) = 483.00 p < .001 ηp2 = .90	F(1,54) = 157.80 p < .001 ηp2 = .75	NS	F(1,54) = 27.29 p < .001 ηp2 = .34	NS	NS	NS
Intelligibility	F(1,54) = 133.00 p < .001 ηp2 = .71	F(1,54) = 41.79 p < .001 ηp2 = .44	NS	F(1,54) = 22.11 p < .001 ηp2 = .29	NS	NS	NS
Cognitive abilities	F(1,54) = 855.40 p < .001 ηp2 = .94	F(1,54) = 773.10 p < .001 ηp2 = .94	F(1,54) = 5.11 p = .028 ηp2 =.09	F(1,54) = 738.20 p < .001 ηp2 = .93	NS	F(1,54) = 5.95 p = .018 ηp2 = .10	F(1,54) = 4.33 p = .042 ηp2 = .07
Speech-language impairment	F(1,54) = 679.80 p < .001 ηp2 = .93	F(1,54) = 152.00 p < .001 ηp2 = .74	NS	F(1,54) = 108.90 p < .001 ηp2 = .67	NS	NS	NS
Likeability	F(1,54) = 30.30 p < .001 ηp2 = .36	F(1,54) = 78.61 p < .001 ηp2 = .59	NS	F(1,54) = 15.17 p < .001 ηp2 = .22	NS	NS	NS
Anxiety	F(1,54) = 37.30 p < .001 ηp2 = .41	F(1,54) = 52.19 p < .001 ηp2 = .49	NS	F(1,54) = 12.86 p < .001 ηp2 = .19	NS	NS	NS

Figure 1.

Age ratings in years as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Speech rate ratings are shown in Figure 2A–B. Similar to the age analysis, there were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. For the child talkers, modifying speech to sound older yielded a slower perceived speech rate (unmodified, M = 2.72; modified, M = 1.96, t(55) = 11.94, p < .001, d = −1.60), while for adult talkers, modifying speech to sound younger yielded a faster perceived speech rate (unmodified, M = 3.56; modified, M = 3.95, t(55) = −6.16, p < .001, d = 0.82). There was no main effect of Awareness or interaction with this factor. This means that making participants aware of the unreliability of voice age characteristics did not mitigate the age-related bias in judgments of speech rate.

Figure 2.

Speech rate ratings as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Fluency ratings are shown in Figure 3A–B. Paralleling age and speech rate analyses, there were main effects of Talker Age and Modification, as well as a reliable interaction between Talker Age and Modification. For both child and adult talkers, age modification of speech yielded lower fluency ratings; however, there was a larger effect of modification on fluency ratings for child talkers (unmodified, M = 4.99; modified, M = 3.83, t(55) = 10.31, p < .001, d = 1.38) than for adult talkers (unmodified, M = 7.30; modified, M = 6.88, t(55) = 6.8, p < .001, d = 0.91). Similar to the speech rate analysis, there was no main effect of Awareness or interactions with Awareness; making participants aware of the unreliability of voice age characteristics did not mitigate effects of the speech modification on perceived speech fluency.

Figure 3.

Fluency ratings as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Intelligibility ratings are shown in Figure 4A–B. As with the other measures, the ANOVA revealed main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. As with fluency ratings, spectral modification of a talker’s speech to sound older or younger yielded lower ratings of intelligibility; however, effects of age modification on intelligibility ratings were significant for child talkers (unmodified, M = 86.2; modified, M = 74.2, t(55) = −11.1, p < .001, d = −1.48), but not for adult talkers (unmodified, M = 97.8; modified, M = 94.8, t(55) = −1.64, p = .1). The lack of a main effect of Awareness or an interaction with Awareness reveals that directing participants’ attention to the unreliability of voice age characteristics did not affect intelligibility judgments, nor did it mitigate the effects of spectral modification on intelligibility judgments.

Figure 4.

Intelligibility ratings as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Percentages of talkers classified with probable cognitive impairment are shown in Figure 5A–B. There were main effects of Talker Age, Modification, and Awareness, as well as Talker Age × Modification and Modification × Awareness interactions, and a three-way between Talker Age × Modification × Awareness interaction. Modifying child talkers’ speech substantially increased cognitive impairment classification (unmodified, M = 11.0%; modified, M = 86.0%, t(55) = −27.92, p < .001, d = −3.73). In contrast, modifying adult talkers’ speech yielded no effect on cognitive impairment classification (unmodified, M = 1.6%; modified, M = 1.6%, t(55) = 0.0, p > .99). The effect of Awareness and interactions with Awareness suggests that making participants aware of the unreliability of age characteristics somewhat lessened the effect of the modification on cognitive impairment assessments of child talkers; however, effects of the modification on assessments of child talkers were reliable for both the unaware (unmodified, M = 11.4%; modified, M = 91.3%, t(32) = −27.9, p < .001, d = −4.86) and the aware condition (unmodified, M = 10.6%; modified, M = 78.4%, t(22) = −14.2, p < .001, d = −2.96).

Figure 5.

Percent judged to have a cognitive impairment as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Percentages of talkers classified with probable speech-language impairment are shown in Figure 6A–B. There were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. Children were much more likely to be judged to have speech-language impairments than adults. Modifying child talkers’ speech produced a large increase in assessed speech-language impairments (unmodified, M = 44.0%; modified, M = 93.5%, t(55) = −12.5, p < .001, d = −1.67), but the modification had no effect on assessed speech-language impairment in adult talkers (unmodified, M = 5.5%; modified, M = 5.5%, t(55) = 0.01, p = .99). The lack of main effect of or interaction with Awareness reveals that directing participants’ attention to the unreliability of voice age did not alter assessments of speech-language impairments. This lack of interaction is striking, given that the age modification more than doubled the rate of assessed speech language impairments for child talkers.

Figure 6.

Percent judged to have a speech-language delay or disorder as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Likeability ratings are shown in Figure 7A–B. As with the other measures, there were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. Spectral modification of the talker’s speech yielded less likable talkers; however, there was a larger effect of Modification on likeability ratings for child talkers (unmodified, M = 2.68; modified, M = 3.62, t(55) = −10.15, p < .001, d = −1.36) than for adult talkers (unmodified, M = 2.38; modified, M = 2.69, t(55) = −2.79, p < .01, d = −0.38). The lack of interaction with the Awareness factor reveals that directing participants’ attention to the unreliability of voice age did not alter perception of the rated attribute.

Figure 7.

Likeability ratings as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Anxiety ratings are shown in Figure 8A–B. There were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. Adults were generally perceived to be less anxious than children, and spectral modification of the talker’s speech yielded greater perceptions of talker anxiety. Unlike many other measures, however, effects of Modification on the rated attribute were larger for adult talkers (unmodified, M = 4.79; modified, M = 4.17, t(55) = 8.64, p < .001, d = 1.15) than child talkers (unmodified, M = 3.86; modified, M = 3.65, t(55) = 2.36, p < .05, d = 0.32). As with most measures, there was no main effect of or interaction with Awareness, indicating that directing participants’ attention to the unreliability of voice age did not alter perceptions of talker anxiety.

Figure 8.

Anxiety level as a function of Talker Age and Modification for participants in (A) the Unaware condition or (B) the Aware condition of Exp. 1a, as well as (C) SLP Master’s student participants in Exp. 1b.

Discussion

Exp. 1a used a spectral manipulation to alter voice age characteristics of talkers to determine its possible effect on judgments of prosody, speech, and talker characteristics. The manipulation was very successful in altering perceived talker age. Modification made child talkers sound much older and adult talkers much younger. Moreover, altering perceived talker age influenced judgments on a wide range of dimensions. Speech of child talkers which was modified to sound like that of older talkers was judged to be slower, less fluent, and less intelligible. Child talkers with modified speech were judged to be less likeable, more anxious, and more likely to have cognitive and/or speech-language impairments. Effects of modification on judgments for adult talkers were less pronounced, but still quite apparent. Speech of adult talkers that was modified to sound like it was produced by much younger talkers was judged to be faster and less fluent, but not less intelligible. Adult talkers with modified speech were judged to be less likeable and more anxious, but not more likely to have cognitive or speech-language impairments.

A second question addressed here was whether awareness of the unreliability of voice age characteristics would reduce bias in judgments of talkers’ speech and attributes. The answer to this question is generally no. Results revealed no reduction in degree of age-related bias for most measures, including prosody-related measures (speech rate, fluency), as well as intelligibility, likelihood of a speech-language disorder or delay, likeability, and anxiety. Only for assessments of cognitive impairment was there a significant difference in degree of age-related bias observed for participants made aware of the unreliability of voice age, relative to the group which was not aware of this issue. Effects of awareness on cognitive abilities ratings were limited to reductions in the frequency of rating as cognitively impaired child talkers in the modified speech condition. Overall, these findings reveal that age-related bias in judgments of prosody, speech, and impairment are relatively automatic, such that the listener has little or no control over such bias, an issue that has implications for developmental prosody research as well as clinical practice. These findings also support the hypothesis that ratings of prosody and other speech attributes depend not only on isolated characteristics of a talker’s speech, but on the listener’s overall assessment of the talker’s characteristics of an individual, including those related to probable impairment.

One question is whether participants may have guessed the purpose of the experiment and were simply responding in a manner that conformed to how they were supposed to respond. To address this question, post-experiment surveys were administered to all participants that asked participants open-ended questions about their impressions of the talkers and files, as well as what they thought the purpose of the experiment was. Results revealed that 0% of participants in either group guessed the purpose, which lessens concern about the role that demand characteristics may have played in these findings. In sum, the results of Exp. 1a show that (a) naïve listeners are influenced by the perceived age of the talker on many dimensions in judging speech and talker characteristics and (b) awareness of the unreliability of voice age characteristics does little to mediate any of observed age bias.

Experiment 1b

In Exp. 1b, we considered whether the main finding of Exp. 1a – namely, that perceived talker age influences judgments of prosodic, speech, and impairment characteristics – also extended to a sample of SLP Master’s students. Because such individuals have had considerably more training and exposure to speech-language issues than the undergraduate psychology participants in Exp. 1a, they might be less susceptible to age-related bias in assessment of prosodic, speech, and impairment characteristics.

Methods

Participants and Design

Participants were 24 SLP Master’s students (all female) with self-reported normal hearing who were native speakers of American English and at least 18 years of age (M = 24.7 years). All participants were enrolled at Michigan State University and received course credit or monetary compensation for participation. Twenty-one of the students were in their first year of the Master’s program, and three were in their second year. All had completed a full sequence of prerequisite courses in communication sciences and disorders, including dedicated courses in language development, speech-language disorders, speech-language evaluation and treatment procedures, speech sciences, hearing sciences, phonetics, and other courses. The experiment implemented a 2 (Talker Age: child, adult) × 2 (Modification: unmodified, modified) within-subject factorial design.

Stimuli, Apparatus, Task, and Procedure

The stimuli and equipment were the same as Exp. 1a. The task and procedure matched all details of the unaware condition in Exp 1a.

Results

A 2 (Talker Age: child, adult) × 2 (Modification: unmodified, modified) repeated measures ANOVA was conducted for each dependent variable; results are summarized in Table 3. Age estimates are shown in Figure 1C. As in Exp. 1a, the modification was highly successful in creating different perceived talker ages. Unmodified speech sounded age-appropriate (child talkers, M = 6.5 years; adult talkers, M = 29.1 years), while modifying speech made child talkers sound older (M = 23.0 years) and adult talkers sound younger (M = 14.6 years), accounting for the significant Talker Age × Modification interaction. Post-hoc paired-samples t-tests showed the modification affected estimated talker age for both child talkers, t(23) = −15.4, p < .001, d = −3.13, and adult talkers, t(23) = 14.5, p < .001, d = 2.96.

Table 3.

Summary of results of two-way ANOVAs for Exp. 1b.

	Main Effect		Interaction
Dependent Variable	Talker Age	Modification	Talker Age × Modification
Age	F(1,22) = 133.99 p < .001, ηp2 = .86	F(1,22) = 5.44 p = .029, ηp2 = .20	F(1,22) = 237.58 p < .001, ηp2 = .92
Speech Rate	F(1,22) = 168.48 p < .001, ηp2 = .88	F(1,22) = 8.34 p = .009, ηp2 = .28	F(1,22) = 61.41 p < .001, ηp2 = .74
Fluency	F(1,22) = 176.45 p < .001, ηp2 = .89	F(1,22) = 73.32 p < .001, ηp2 = .77	F(1,22) = 7.51 p = .012, ηp2 = .25
Intelligibility	F(1,22) = 59.04 p < .001, ηp2 = .73	F(1,22) = 56.36 p < .001, ηp2 = .72	F(1,22) = 34.98 p < .001, ηp2 = .61
Cognitive abilities	F(1,22) = 243.66 p < .001, ηp2 = .92	F(1,22) = 198.81 p < .001, ηp2 = .90	F(1,22) = 189.02 p < .001, ηp2 = .90
Speech-language impairment	F(1,22) = 315.05 p < .001, ηp2 = .94	F(1,22) = 46.88 p < .001, ηp2 = .68	F(1,22) = 33.53 p < .001, ηp2 = .60
Likeability	F(1,22) = 16.93 p < .001, ηp2 = .44	F(1,22) = 39.90 p < .001, ηp2 = .65	NS
Anxiety	F(1,22) = 16.08 p = .001, ηp2 = .42	F(1,22) = 16.98 p < .001, ηp2 = .44	F(1,22) = 7.83 p = .010, ηp2 = .26

Speech rate assessments are shown in Figure 2C. Similar to Exp. 1a, there were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. Modifying child talkers’ speech to sound older yielded slower perceived speech rates (unmodified, M = 2.98; modified, M = 2.07, t(23) = 10.75, p < .001, d = −2.19), while modifying adult talkers’ speech to sound younger yielded faster perceived speech rates (unmodified, M = 3.40; modified, M = 3.91, t(23) = −5.60, p < .001, d = 1.14). Results show that SLP Master’s students were susceptible to the same age-related bias in judging speech rate as the naïve participants tested in Exp. 1a.

Fluency ratings are shown in Figure 3C. Similar to Exp. 1a, there were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. Modification yielded lower ratings of fluency for both child and adult talkers, with a larger effect on fluency ratings for child talkers (unmodified, M = 5.64; modified, M = 4.60, t(23) = 8.65, p < .001, d = 1.76) than for adult talkers (unmodified, M = 7.26; modified, M = 7.04, t(23) = 2.39, p < .05, d = 0.49).

Intelligibility ratings are shown in Figure 4C. Similar to Exp. 1a, there were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. As with fluency ratings, modification yielded lower intelligibility ratings for both child and adult talkers. Modification effects were significant for both child talkers (unmodified, M = 91.9; modified, M = 80.6, t(23) = −14.2, p < .001, d = −1.66) and adult talkers (unmodified, M = 99.4; modified, M = 97.8, t(23) = −2.99, p < .01, d = −0.61), with the Talker Age × Modification interaction due to the stronger effect of modification on ratings for child than adult talkers. These findings suggest that SLP Master’s students are strongly biased to interpret the same set of articulatory cues produced by children as less intelligible when those cues occurred with an adult’s voice (modified condition) than with a child’s voice (unmodified condition).

Percentages of talkers classified as cognitively impaired are shown in Figure 5C. Similar to Exp. 1a, there were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. Child talkers were more likely to be classified as having a cognitive impairment than adults. Moreover, modifying child talkers’ speech produced a large increase in classification of talkers as having a cognitive impairment (unmodified, M = 5.6%; modified, M = 83.8%, t(23) = −14.1, p < .001, d = −2.88), while modifying adult talkers’ speech had no effect on whether they received a cognitive impairment classification (unmodified, M = 1.9%; modified, M = 1.4%, t(23) = 0.30, p = .77).

Percentages of talkers judged to have a speech-language impairment are shown in Figure 6C. Similar to most other measures, there were main effects of Talker Age and Modification, as well as a significant Talker Age × Modification interaction. Child talkers were more likely than adults to be classified as having a speech-language impairment; modifying child talkers’ speech almost doubled the likelihood of classifying a talker as having a speech-language impairment (unmodified, M = 50.8%; modified, M = 97.5%, t(23) = −9.5, p < .001, d = −1.94). Modification also had a significant, though weaker, effect on percentages of adult talkers classified with a speech-language impairment (unmodified, M = 3.2%; modified, M = 11.6%, t(23) = −2.58, p < .05, d = −0.53), accounting for the Talker Age × Modification interaction.

Likeability ratings are shown in Figure 7C. Similar to Exp. 1a, there were main effects of Talker Age and Modification, as well as a reliable Talker Age × Modification interaction. As in Exp. 1a, modification of speech generally produced less likable talkers. The Talker Age × Modification interaction revealed that the effect of Modification on likeability was reliable for child talkers (unmodified, M = 2.19; modified, M = 3.14, t(23) = −6.26, p < .001, d = −1.26), but not adult talkers (unmodified, M = 2.18; modified, M = 2.33, t(23) = −1.07, p = .29).

Anxiety ratings are shown in Figure 8C. There were main effects of Talker Age and Modification, but there was no reliable Talker Age × Modification interaction. Adults were judged as less anxious than children, and modification increased perceptions of talker anxiety. Although the Talker Age × Modification interaction was not reliable, post-hoc t-tests revealed a significant effect of Modification for adult talkers (unmodified, M = 4.73; modified, M = 4.26, t(23) = 4.06, p < .001, d = 0.83), but not child talkers (unmodified, M = 4.08; modified, M = 3.93, t(23) = 0.99, p = .33).

Discussion

Exp. 1b replicates the general pattern of findings observed in Exp. 1a in showing that SLP Master’s students are also susceptible to similar age-related bias in assessing talker characteristics on a range of dimensions. As in Exp. 1a, when the speech of child talkers was spectrally modified to sound as if the talkers were older, it was judged to be slower, less fluent, and less intelligible than unmodified child speech. Child talkers with spectrally-modified speech were also judged to be less likeable and were more likely to be judged to have cognitive and/or speech-language impairments. Also similar to Exp. 1a, effects of age-modified adult speech on judgments of talker characteristics were less pronounced, but still quite apparent. Modified speech of adult talkers was judged to be faster and less fluent, but not less intelligible. Adult talkers with age-modified speech were also judged to be less likeable and more anxious, and were also more frequently judged to have speech-language impairments. The spectral modification, however, did not impact judgments of whether adult talkers had a cognitive impairment.

In sum, Exp. 1b showed that the specialized speech and language training and experience of the SLP Master’s students does not inoculate them against age bias in judgments of talker characteristics that are largely predicted to be unrelated to talker age. These findings indicate that age-related bias may be difficult to overcome even with additional training in speech-language issues. However, experience with and exposure to speech-language issues and language development did not appear to create more age-related bias in the SLP Master’s students’ judgments compared with those of the undergraduates of Exp. 1a, as would have been predicted by findings that speech perception can be affected by experience with and exposure to particular talker groups or group characteristics (e.g., Drager, 2011). Overall, the findings support the hypothesis that ratings of prosody and other speech attributes depend not only on isolated characteristics of a talker’s speech, but on the listener’s overall assessment of an individual.

One question raised here is the extent to which intelligibility rating differences across spectral modification conditions in Exps. 1a–1b could have been due to signal degradation associated with the spectral modification itself, as opposed to age-related bias. This possibility was considered in Exp. 2.

Experiment 2

Exp. 2 presented a new sample of naïve participants with a subset of the child talker and adult talker speech in unmodified and modified conditions, and participants were asked to transcribe the words they heard. If the intelligibility effects observed in Exp. 1a and 1b were due simply to signal degradation associated with the spectral modification, then transcription accuracy would be expected to be worse for the modified speech stimuli than for the unmodified speech stimuli. On the other hand, if the intelligibility effects observed in Exp. 1a and 1b reflected an age bias, then transcription accuracy in Exp. 2 would be expected to be similar in the modified and unmodified talker conditions.

Methods

Participants and Design

Participants were 16 undergraduate students (7 F, 9 M) with a mean age of 20.8 who received course credit for their participation. A 2×2 within-subjects factorial design was used with independent variables of Talker Age (adult, child) and Modification (unmodified, modified). The dependent variable was the proportion of words correctly transcribed by participants in each condition.

Stimuli

Stimuli for Exp. 2 were a subset of the stimuli described in Exp. 1; in particular, the speech fragment with the largest word count for each talker was selected. Mean word count for child and adult files was 43.7 words and 65.1 words, respectively, and the mean duration of child and adult files was 27.0 sec and 27.7 sec, respectively.

Apparatus and Procedure

Lenovo Intel® Core™2 Duo CPU E8500 computers with 19-inch screens running Microsoft Word 2007 and Microsoft Media Player 10 connected to Sennheiser HD 280 Professional headphones were used for the experiment. Two counterbalanced lists of speech files were created, each of which consisted of 18 speech fragments (one from each of the nine child and nine mother talkers). Half of the fragments were in the unmodified condition and half in the modified condition. Each speech fragment occurred only once in the entire list in one type of modification. Participants were randomly assigned to one of the two lists, with equal numbers assigned to each list. Participants were directed to the folder on the computer which contained the stimuli for their list, and instructed to listen to the files in the order specified on the computer screen. Participants were asked to listen to a speech file in Microsoft Media Player as many times as necessary and then to transcribe the speech of the main speaker in each file into a template in a Microsoft Word document. They were instructed to use only lowercase letters in their transcription and not to use punctuation to facilitate comparison of transcriptions. In addition, they were instructed not to transcribe any non-speech noises (e.g. laughter, singing). Participants were also asked to transcribe partial words with the heard phonemes followed by a dash (e.g. he-, wh-), to transcribe filler words as “um” and “uh”, and to transcribe “X” for all consecutive unintelligible words. The entire experiment took 60–90 minutes.

Data Analysis

A transcription key was first prepared for each speech fragment by two trained phonetic analysts from the acoustic-phonetic information in the fragment’s unmodified version. Of particular interest was identifying words in each fragment which were unanimously deemed intelligible by both of the trained phonetic analysts, as well as those words in the fragment which were unintelligible for reasons other than spectral modification, e.g., developmental misarticulation, soft speech, background noise, etc. Our goal was to compare transcription accuracy by naïve listeners in unmodified vs. modified speech conditions to words marked unanimously intelligible to the trained phonetic analysts in the unmodified speech fragment, which made it possible to determine the extent to which spectral modification per se degraded intelligibility. Phonetic analysts were given the same guidelines for transcription as participants, but listened to the unmodified speech files while viewing spectrograms and waveforms in Praat (Boersma & Weenink, 2002). The use of combined acoustic and auditory cues by the trained phonetic analysts was expected to generate excellent agreement between the two analysts on words which were intelligible in the unmodified speech files. The transcription key for each file consisted of words that were transcribed by (and thus intelligible to) both analysts; words which were not agreed on exactly by both analysts and those transcribed as “X” were treated as unintelligible. The transcription key was then compared against the naïve listeners’ transcripts to determine the rate with which naïve participants correctly identified each word in the unmodified vs. modified conditions. This permitted an estimate of the extent to which the spectral modification might have reduced intelligibility of the speech. This metric was calculated for each participant for each Talker Age and Modification condition by dividing the number of words correctly identified by the total number of words identified in the transcription key for that file.

Results

The proportion of intelligible words in each condition are shown in Figure 9. The mean of intelligible words heard for the child speech was 76.0% for unmodified, 70.6% for modified, and for the adult speech was 92.6% for unmodified, 92.9% for modified. A 2 (Talker Age: child, adult) × 2 (Modification: unmodified, modified) ANOVA by subjects was conducted on this data; both Talker Age and Modification were within-subject factors. There was a significant effect of Talker Age, F(1,15) = 791.50, p < .001, η_p² = .98, but no effect of Modification, F(1,15) = 1.69, p = .214, and no interaction, F(1,15) = 1.89, p = .190.

Figure 9.

Average of transcribed intelligible words as a function of Talker Age and Modification for Exp. 2.

Discussion

The main finding from this study was that the spectral modification employed here did not significantly reduce intelligibility in a transcription task. In particular, the lack of a significant effect of Modification and absence of an interaction with Modification indicates that the modification was not responsible for degradation in the proportion of words understood. These results suggest that the effects of Modification on perceived intelligibility observed in Exps. 1a–1b were primarily the result of age-related bias as opposed to actual differences in intelligibility per se between modified and unmodified speech. The findings of Exp. 2 therefore lend support to our interpretation from Exps. 1a–1b that listeners were biased to interpret the same articulatory cues produced by children as less intelligible when those cues occurred with an adult’s voice (child-modified condition) than when they occurred with a child’s voice (child-unmodified condition), since misarticulations are common for 5-year-old children but rare for typical adults. Returning to the issue of prosody, we asked in Exp. 3 whether a different method of reporting prosody might result in less age-related bias than the impressionistic judgments of Exps. 1a–1b.

Experiment 3

Methods

Participants and Design

Participants were 10 prosody analysts (8 F, 2 M) at Michigan State University (mean age 20.6 years) who volunteered or received pay for participation. All analysts had participated in a one-semester seminar on the Rhythm and Pitch (RaP) system of prosody transcription (Breen, Dilley, Kraemer, & Gibson, 2012; Dilley & Brown, 2005), which is a method of annotating prosody in speech, including phrasal boundaries, prominences, pitch accents, disfluencies, and other information. The seminar was taught by the first author; participants passed tests of proficiency at the end of the seminar. The seminar included weekly practice in transcription, and participants attended weekly group transcription practice sessions following the seminar. A 2×2 within-subjects factorial design was used with independent variables of Talker Age (adult, child) and Modification (unmodified, modified). The dependent variables were the number of phrasal boundaries and disfluencies marked by analysts in the four conditions (adult-unmodified, adult-modified, child-unmodified, child-modified).

Stimuli

Exp. 3 materials included the exhaustive stimulus set described in Exp. 1a. There were 105 unmodified fragments from the 18 talkers (9 child, 9 adult) and their 105 modified counterparts.

Procedure

The analysts used the Rhythm and Pitch (RaP) labeling system to annotate the locations and types of disfluencies and phrasal boundaries. Prior to beginning work on the project analysts reviewed the different types of disfluencies, listening to multiple examples of each type. Next, analysts were allowed to listen to a speech file as many times as necessary before providing an annotation. Two counterbalanced lists of speech files were created so that no analyst would ever listen to the same speech file in both modification conditions. Within a list, the sequential pairing of speech files with Modification conditions was random but fixed for all participants assigned to that list. The analysts were instructed to listen to files in a designated folder sequentially from the first to the last file. Half the fragments on each list were from child talkers and half were from adult talkers; in addition, half of the fragments were in the unmodified condition and half in the modified condition. This counterbalancing ensured that each analyst never heard the same speech file in both its unmodified and modified form.

Analysts were randomly assigned to one of the two speech file lists (five analysts per list). They labeled the phrasal boundaries and disfluencies using a modified version of RaP guidelines (Dilley & Brown, 2005) which involved indicating when a prosodic phrase boundary occurred and how large it was. In particular, labels of “)?”, “)”, “))?”, and “))” were used to indicate a possible minor phrasal boundary, a clear minor phrasal boundary, a possible major phrasal boundary, or a clear major phrasal boundary, respectively. Finally, analysts annotated each time a disfluency was perceived to occur.

Apparatus

The computers used for prosody analysis were either Lenovo Intel® Core™2 Duo CPU E8500 with 19 inch screens using Sennheiser HD 280 Professional headphones from the Michigan State University Speech Lab or personal computers owned by the analysts.

Analysis

The rate of marking a phrasal boundary [“)”, “))?”, or “))”] was first calculated for each labeler in each condition, which corresponded to the number of times a phrasal boundary was marked, divided by the number of possible phrasal boundary locations (i.e., the number of word-final locations in the file, calculated as the number of words in the file minus one, since the final word in the file was trivially a phrasal boundary; Breen et al., 2012). The rate of marking a disfluency corresponded to the number of times a disfluency was marked by each labeler in each condition, divided by the number of possible disfluency locations (i.e., the number of words or pauses minus one - see above).

Analyses of inter-rater agreement were also carried out using the kappa (κ) metric, which adjusts for chance agreement levels based on the number of coding distinctions and label frequency (Carletta, 1996). κ is calculated as in (1); A_E is expected chance agreement and A_O is observed (actual) agreement:

$$\kappa =\frac{A_O-A_E}{1-A_E}$$ (1)

Values of κ were calculated following Breen et al. (2012). Here, specific labels were grouped into label equivalence relations to indicate how labels corresponded to the constructs of interest (i.e., presence and size of phrasal boundary and presence of disfluency). The label equivalence relations for this study are shown in Table 4. See Breen et al. (2012) for further details of calculation of κ.

Table 4.

Label equivalence relations used for calculating rates of labeling boundaries and disfluencies for Exp. 3.

Presence of phrasal boundary	Boundary	No boundary
	), ))?, ))	)?, no label
Size of boundary	Large	Small
	))?, ))	)?, ), no label
Disfluency vs. No disfluency	Disfluency	No-disfluency
	prolongation, truncation, repeat, mispronunciation, unfilled pause	restart, no label

Results

A 2 × 2 repeated measures ANOVA was conducted by-subjects with Talker Age and Modification as within-subject factors for each dependent variable. The proportion of phrasal boundaries marked is shown in Figure 10. More phrasal boundaries were identified for child speech (M = 0.39) than adult speech (M = 0.21). There was a significant effect of Talker Age, F(1,9) = 304.73, p < .001, η_p² = .97, but no effect of Modification, F(1,9) = 0.44, p = .52, and no interaction, F(1,9) = 0.53, p = .49. Values of κ ranged from 0.74 to 0.79, indicating high and substantial agreement.

Figure 10.

Proportion of prosodic phrase boundaries marked by labelers as a function of Talker Age and Modification for Exp. 3.

The proportion of large phrasal boundaries marked is shown in Figure 11. More large phrasal boundaries were identified for child speech (M = 0.31) than adult speech (M = 0.14). There was a significant effect of Talker Age, F(1,9) = 142.56, p < .001, η_p² = .94, but no significant effect of Modification, F(1,9) = 1.87, p = .204, and no interaction, F(1,9) = 0.06, p = .81. Values of κ ranged from 0.82 to 0.86, indicating high and substantial agreement.

Figure 11.

Proportion of large prosodic phrase boundaries marked by labelers as a function of Talker Age and Modification for Exp. 3.

The proportion of disfluencies marked is shown in Figure 12. There was a significant effect of Talker Age, F(1,9) = 140.09, p < .001, η_p² = .94; more disfluencies were identified for children’s speech (M = 0.14) than adult speech (M = 0.02). Moreover, there was a marginally significant effect of the Modification on the rate of labeling disfluencies, F(1,9) = 4.77, p = .057, as well as a marginally significant interaction, F(1,9) = 3.45, p = .096. Values of inter-rater agreement across conditions ranged from κ = .26 to .38, indicating moderate agreement.

Figure 12.

Proportion of disfluencies marked by labelers as a function of Talker Age and Modification for Exp. 3.

Discussion

In this experiment, analysts used a formal prosody labeling system to indicate prosodic phrase boundaries and disfluencies in speech fragments produced by children and adults, where half of the fragments had been modified to make talkers sound older or younger, respectively. Careful counterbalancing ensured that no analyst heard any speech fragment more than once in a single modification condition. Analysts marked more prosodic phrase boundaries and disfluencies in children’s speech compared with adults’ speech. Critically, spectral modification did not significantly affect rates of indicating prosodic phrase boundaries or their perceived sizes. There was also no significant effect of the spectral manipulation on the rate of labeling disfluencies, although the effect approached significance. It is notable that effect sizes of the modification on rates of labeling disfluency were uniformly smaller than effect sizes of the manipulation on impressionistic fluency ratings in Exp. 1. Results suggest that using a formal prosodic labeling system incurs less age-related bias in judging prosody than impressionistic judgments of prosody. This supports the usage of formal prosody labeling as a particularly valuable methodology for investigation of developmental prosody issues while minimizing age-related bias.

General Discussion

The present research used spectral manipulation to alter talkers’ perceived ages and determine the subsequent effects on judgments of prosodic, segmental, clinical, and personal characteristics of speech and/or the talkers themselves. Understanding the effects of perceived age on talker and/or speech judgments is important for developmental work on speech and language. The extent to which talker age might affect such judgments was investigated in experiments using impressionistic ratings (Exp. 1a and 1b), orthographic transcription (Exp. 2), and formal prosody labeling (Exp. 3).

Exp. 1 demonstrated that differences in perceived talker age resulting from our spectral manipulation were associated with differences in every impressionistic measure investigated, including prosodic and articulatory variables (perceived speech rate, fluency, intelligibility), and variables related to talkers themselves (likelihood of speech-language and/or cognitive impairment, likeability, and anxiety). This is the first study to demonstrate that manipulations to F0 and formant frequencies affect judgments of talker age for children’s speech; such effects have previously been demonstrated for adults’ speech (Harnsberger et al., 2008; Reubold et al., 2010). Moreover, Exp. 2 confirmed that effects of perceived age on intelligibility ratings could not be attributed to signal degradation due to spectral modification. Finally, Exp. 3 showed that age-related bias in judgments of at least some prosodic variables may be mitigated by using a formal prosodic coding system. Overall, these findings confirm that perceived talker age affected listeners’ standards for judging speech and talker attributes, demonstrating the existence of age-related bias in impressionistic judgments of these speech and talker characteristics.

Exp. 1a investigated whether the age-related bias in judgments of speech and talker characteristics can be reduced as a function of attention, or whether such bias is automatic. This was accomplished by comparing ratings by undergraduate students made aware of the spectral manipulation with ratings by naïve undergraduates. No reduction in degree of age-related bias was observed for the former group for the majority of measures. These findings suggest that age-related bias is relatively automatic.

Additionally, investigating ratings in Exp. 1b by SLP Master’s students allowed us to determine whether additional learning and experience relating to speech-language issues affected age-related bias in judgments of prosody, speech, and talker characteristics. While the majority of measures showed no differences in age-related bias across modification conditions, three measures showed a reduction in bias for SLP Master’s students relative to undergraduates: age estimates, fluency judgments, and likeability. These findings suggest that additional learning and experience in speech-language issues may reduce age-related bias for at least some types of judgments, but that in general, age-related bias in judgments of prosodic, segmental, and impairment characteristics may be difficult to overcome. It is possible that had our participants been certified SLPs, who have had even more training and experience, even less age-related bias might have been observed, an issue which remains for future research.

It is worth considering whether an alternative explanation exists for effects of the spectral manipulation on judgments according to which listeners detected incongruity between the talker’s true age and his/her perceived age and judged such “mismatched” speech more negatively than “matched” speech. Such an explanation predicts more negative judgments for “mismatched” speech on relevant dimensions than “matched” speech. In contrast to this prediction, for a number of properties for which one scale endpoint could be interpreted as more negative, i.e., cognitive impairment, language impairment, and intelligibility, “mismatched” adult speech was not assessed by listeners more negatively than “matched” speech by most listener groups across Exps. 1a–1b. These findings argue against an explanation of the spectral modification manipulation effects based solely around incongruity detection. The age bias explanation is most consistent with the majority of results across these experiments; however, we can’t rule out that incongruity detection did not play a role in some conditions examined.

Our results suggest that age-related voice characteristics can affect a wide range of impressionistic judgments of attributes which are commonly assumed to be orthogonal or quasi-orthogonal to prosodic, segmental, and talker attributes. These findings are consistent with prior research demonstrating the interdependence of perceptual judgments of pitch, timing, and/or spectral profile (Bond et al., 1988) as well as findings showing that age estimates are affected by manipulations to F0, F1, and/or speech rate (Harnsberger et al., 2008; Reubold et al., 2010). In terms of theory, the present research suggests that prosody cannot be compartmentalized from other components of the speech signal (e.g., the segmental and indexical components, where the latter includes talker age). Our findings therefore are consistent with a wide range of research demonstrating interdependency in perceiving and processing lexico-segmental, prosodic, and indexical components of speech (e.g., Ladefoged & Broadbent, 1957; Nygaard, Sommers, & Pisoni, 1994).

As expected, measures of speech and talker characteristics were also substantially affected by the baseline speech-language competency of talkers. Across studies, children were judged as being less competent (e.g., less fluent, slower in rate, less intelligible, more likely to be impaired) than adults. More interesting is that baseline speech-language competency interacted with spectral modification, such that children’s speech in general was judged very differently on many measures, depending on voice age characteristics, while adults’ speech was less often and less substantially affected by the modification.

One consistent finding was that when children’s speech was spectrally modified, talkers were particularly likely to be interpreted as having speech-language and cognitive impairments and to be significantly less likeable. These findings can be explained by the pairing of relatively high disfluency and misarticulation rates, limited vocabulary, and simpler syntax of children’s speech, on the one hand, with an adult-sounding voice, on the other; such a situation is consistent with an adult with below-par speech-language and cognitive skills. Given that adult talkers with communication impairments are judged more negatively than talkers without such impairments (Franck et al., 2003; Lass et al., 1993), we can explain the decreased likeability of child talkers in the modified speech condition. The findings support the hypothesis that prosody and other speech ratings depend not only on perceived speech characteristics, but on an assessment of the talker as an individual, including possible impairment. Overall, the present research illustrates the potential of spectral modification as a tool for investigations of attitudes toward groups in relation to developmental, lifespan, and/or communication impairment issues.

Moreover, these findings raise cautionary notes for developmental research studies in which voice age characteristics covary with participant groups under study, or where sufficient time elapses during the observation phase that maturation of voice age occurs. Our findings suggest that confounding may occur between primary prosodic variables of interest (e.g., intelligibility, speech rate) and voice age characteristics in longitudinal or cross-sectional developmental research designs, an issue that has seldom considered in developmental studies of prosody. Many prosodic characteristics have been investigated in child speech, including phrasal boundary emergence and placement (Crystal, 1986), timing (Snow, 1997), disfluency (Gordon & Luper, 1989; Wexler & Mysak, 1982), speaking rate (Kelly & Conture, 1992) and phrase-level intonation (Chen, 2011; MacWhinney & Bates, 1978). In much of this work, children’s prosodic behavior has been compared to that of adults or of children of other ages. This research suggests that in such comparative designs, differences in prosodic metrics could have partly reflected the influence of voice age characteristics, rather than merely the prosodic variables of interest. It would be wise for developmental speech-language research studies to include additional controls to guard against potential confounding between voice age characteristics and primary perceptual speech-related variables of interest if utilizing impressionistic judgments, or else to use more formalized prosodic coding methods.

The present results also have implications for assessment of speech in clinical situations. First, little research has examined the possibility of bias affecting clinical assessment of children’s speech. Existing work suggests that a child’s presumed age and gender influence adults’ perceptions of children’s speech and ratings of its accuracy and quality (Munson et al., 2010; Munson & Seppanen, 2009). These results suggest that a child’s perceived age can bias perceptual judgments of speech in more far-reaching ways. Second, these results suggest that maturation of the voice could occur during clinical observations over time, which could be confounded with perceptions of other speech-language attributes of greater clinical interest (e.g., speech rate, fluency, intelligibility). Third, these results highlight a potential risk of unreliability in impressionistic perceptual judgments of speech variables of interest (e.g., speech rate, fluency, or intelligibility), due to potential misattribution of changes from voice age characteristics.

In addition to these cautionary notes, the present results suggest some positive new directions for therapeutic strategies aimed at altering how a talker’s speech, degree of impairment, and/or self-presentation may be perceived by others. In particular, future research might explore whether some talkers might be coached to change their voice spectral profile or pitch to make themselves sound older or younger in certain situations in order to change others’ perceptions of their speech rate, fluency, intelligibility, anxiety level, or the likelihood that they have cognitive or communicative impairments. For example, future work might explore the value of an adult with a communication impairment adopting something like a “Bart Simpson voice” (i.e., one which sounds more child-like) when talking over the telephone (e.g., Guo & Togher, 2008) in order to alter how others perceive her speech, impairment, or presentational characteristics. Candidates for such an intervention might be individuals who have difficulty controlling aspects of speech articulation or rate but who have good control over voice characteristics, e.g., some dysarthric individuals or people who stutter.

In summary, age-related spectral voice characteristics can affect judgments of speech and talker characteristics in somewhat surprising ways, given that these characteristics are often treated in theory and practice as orthogonal to one another. This research suggests cautionary notes for research and clinical settings when observing individuals of different ages or the same individual over time, since voice age characteristics may be confounded with judgments of prosodic and segmental aspects of speech. Using a formalized coding system for prosody may minimize age-related bias relative to impressionistic ratings. Finally, this work points to potential for developing intervention strategies for talkers with some types of communication impairment that target alterations in voice spectral characteristics to achieve changes in how others perceive the talker’s speech, impairment, and/or self-presentation characteristics.