Implicit Imitation of Regional Dialects in Typically Developing Adults and Adults with. High-Functioning Autism. Thesis

Transcription

1 Implicit Imitation of Regional Dialects in Typically Developing Adults and Adults with High-Functioning Autism Thesis Presented in Partial Fulfillment of the Requirements for the Degree Master of Arts in the Graduate School of The Ohio State University By Sara Phillips-Bourass, B.A. Graduate Program in Linguistics The Ohio State University 2012 Thesis Committee: Professor Cynthia G. Clopper, Advisor Professor Shari R. Speer

2 Copyright Sara Phillips-Bourass, 2012

3 Abstract This study examined the effect of regional dialect on implicit spontaneous phonetic imitation in typically developing adults (TD) and adults with high functioning autism (HFA). Participants shadowed CVC target words produced by talkers from two American English dialect regions, the North and the Midland. Imitation was measured acoustically in terms of vowel quality, vowel duration, midpoint f0, f0 trajectory, and onset and coda consonant duration. It was hypothesized that TD shadowers would imitate familiar dialects less than unfamiliar dialects, based on previous results showing that high frequency words are imitated less than low frequency words (Goldinger, 1998). By extension, frequently encountered or familiar dialects should be imitated less than unfamiliar dialects. Effects of dialect familiarity were in the expected direction, with more imitation of unfamiliar dialects. TD shadowers also imitated perceptually salient targets, including those with long vowels, high f0, or rising f0. The TD productions and targets from the shadowing task were then used as stimulus materials in an AXB discrimination task to measure perceived imitation. Results suggest that TD shadowers were imitating some dialect features. Among the acoustic measures, perceived imitation was correlated with vowel duration and coda duration. HFA shadowers were expected to exhibit less imitation than TD shadowers based on research showing that impaired imitation is a characteristic of HFA (Stieglitz et al., 2008; Stone et al., 1990). This prediction was supported for vowel quality and onset duration, but HFA shadowers imitated vowel duration and midpoint f0 more than TD shadowers. HFA shadowers were less likely than TD shadowers to imitate perceptually long or high f0 vowels, but more likely to imitate long consonants. Effects of dialect familiarity for HFA shadowers were in the opposite direction from those for TD shadowers, with HFA shadowers imitating Northern targets less than the more familiar Midland targets. These ii

4 results suggest an impaired ability for individuals with HFA in imitating unfamiliar or extreme phonetic variation. iii

5 Acknowledgements I would like to thank Cynthia G. Clopper, Shari R. Speer, Kathryn Campbell-Kibler, Mark Pitt, and Jane Stuart-Smith for their insightful comments as well as Kristin L. Rohrbeck and Laura Wagner for help in HFA data collection. I also acknowledge the help of Lauren Eskin, Joselyn Gilbert, and Kenney Hensley in data collection and annotation. Thanks are also due to the audiences at the OSU Cogfest 2010, MCWOP 2010 and the 161st meeting of the ASA as well as the OSU Linguistics Department historical and sociolinguistics discussion group, phonetics and phonology discussion group and psycholinguistics lab group for their feedback. iv

6 Vita B.A. Linguistics, French, Near Eastern Languages and Cultures, Indiana University 2009 to Graduate Fellow, Department of Linguistics, The Ohio State University 2011 to FLAS Fellow, The Ohio State University Publications Phillips, S. and Clopper, C. G. (2012) Perceived imitation of regional dialects. Proceedings of Meetings on Acoustics, Fields of Study Major Field: Linguistics v

7 Table of Contents Abstract Acknowledgements Vita ii iv v List of Figures viii List of Tables ix 1 Introduction Background Approaches to Imitation Imitation and Dialect Frequency and Regional Dialect Dialect Processing Regional Dialect Imitation and Autism Imitation and Autism Dialect and Autism Imitation in Typically Developing Adults Experiment 1: TD Word Shadowing Methods Measures Statistical Models Predictions Results Discussion Experiment 2: AXB Classification Task Methods Predictions Results Discussion vi

8 3 Imitation in Adults with High Functioning Autism Experiment 3: HFA Word Shadowing Methods Measures Statistical Models Predictions Results Discussion General Discussion and Conclusions TD Shadowers HFA Shadowers References A Word List B Linear Mixed-Effects Models: TD Shadowers B.1 Vowel Quality B.2 Onset Consonant Duration B.3 Vowel Duration B.4 Coda Consonant Duration B.5 Midpoint F B.6 F0 Trajectory B.7 Response Time C Linear Mixed-Effects Models: HFA Shadowers C.1 Vowel Quality C.2 Onset Consonant Duration C.3 Vowel Duration C.4 Coda Consonant Duration C.5 Midpoint F C.6 F0 Trajectory C.7 Response Time vii

9 List of Figures 1.1 Vowel plot of the Northern Cities Shift Map of the Northern and Midland dialect regions Vowel plot for Midland and Northern target vowels Vowel plot of baseline, shadowed and target productions, TD shadowers Acoustic imitation score for vowel duration plotted against the duration of the target vowel Correlation between acoustic imitation score and perceived imitation score Vowel plot of baseline, shadowed and target productions, HFA shadowers. 50 viii

10 List of Tables 2.1 Participant groups in Experiment Residential history for Mobile TD shadowers Vowel duration, all TD shadowers Coda consonant duration, all TD shadowers Midpoint f0, all TD shadowers F0 trajectory, all TD shadowers Demographic information for AXB listeners Conditions in Experiment 2 (AXB Classification) Comparing AXB perceived imitation scores to chance Differences between Northern and Midland targets Participant groups in Experiment Residential history for Mobile HFA shadowers Onset duration, comparing HFA and TD shadowers Vowel duration, comparing HFA and TD shadowers Coda duration, HFA shadowers Midpoint f0, comparing HFA and TD shadowers Response time, comparing HFA and TD shadowers Response time, HFA shadowers ix

11 Chapter 1 Introduction It is well known that speakers change their style of speaking depending on the speech of their interlocutor. Conversation partners are known to imitate one another s speech rate, fundamental frequency (Giles et al., 1991; Jungers et al., 2002), voice onset time (Shockley et al., 2004; Nielsen, 2011), and vowel quality (Babel, 2009a; Tilsen, 2009), to name but a few aspects of speech that have been documented. While imitation of speech is an integral part of the way typically developing speakers interact, imitation of gestures is known to be impaired in individuals with high-functioning autism (HFA) (Stieglitz Ham et al., 2008). This study investigates whether imitation of speech is similarly impaired in adults with HFA as compared to typically developing (TD) adults using a shadowing task similar to that in Goldinger (1998). More specifically, this study focuses on regional dialect as a way to investigate the relationship between social information and imitation. Distinctions in regional dialect involve both social information and fine-grained phonetic detail. Autism is characterized by social and linguistic deficits (American Psychiatric Association, 1994), but autistic listeners have unimpaired speech perception and attention to phonetic detail (Mottron et al., 2006; Happé & Frith, 2006). For example, Clopper et al. (2012) reported that adults with HFA are similar to TD adults in terms of their ability to perceive the differences between dialect groups but exhibit an impaired ability to assign stereotyped social information about language attitudes to those same groups. If the degree of imitation is partially due to the influence of sociolinguistic attitudes, HFA shadowers should show different patterns of imitation from 1

12 TD shadowers. Where TD shadowers increase or decrease imitation of a particular dialect group, HFA shadowers should not show as large a difference between target dialects. On the other hand, social factors are not the only influences on imitation. Goldinger (1998) reported that low frequency words were imitated more than high frequency words in a word shadowing task. The present study tested whether this finding also holds true for frequency of exposure to regional dialects. According to the exemplar theories articulated by Goldinger (1998) and Pierrehumbert (2002), targets in an unfamiliar dialect should be imitated more than those in a familiar dialect because an unfamiliar variant has few exemplars stored in memory. Recent perception of that variant should thus exert more influence on subsequent productions than perception of a frequent variant. Such an effect should not differ between HFA and TD shadowers, assuming that the mental representations of fine-grained phonetic detail are the same for HFA and TD populations. 1.1 Background Approaches to Imitation Accommodation Communication Accommodation Theory (CAT, formerly Speech Accommodation Theory) emerged from the observation by Giles (1973) and others that speakers adapt to the speech of their conversation partners depending on social factors. CAT proposes that speakers shift speaking style based on cues from the interviewer, tailoring their speech to that of other speakers rather than to some abstracted idea of formal versus informal situations, as was proposed by Labov (1972). Bell (1984) elaborated a related theory, Audience Design, defining style as speakers response to their audience (p. 145), and claimed that other abstract macro-level influences on speech style such as setting or topic could be explained by their association with different addressees. 2

13 According to CAT, speakers converge or diverge from their interlocutor s speech style for social reasons. They may adopt similar styles, maintain their normal way of speaking, or adopt complementary opposing styles (e.g. Shepard et al., 2001). Speakers may converge when they have positive feelings toward their conversation partner, or if they want to gain their partner s social approval. Divergence, and often maintenance, increases social distance between the interlocutors. Motivations for accommodation can also include improving the effectiveness of communication or expressing group membership (Giles et al., 1991). Crucially for the purposes of our discussion here, the social psychology approach to accommodation focuses on an interactive situation between two speakers (Stuart-Smith, 2011), and often assumes a certain degree of awareness or volition on the part of the participants. In so far as social skills are impaired for speakers with HFA, the explanations given by CAT point to different patterns of imitation for HFA shadowers as compared with TD shadowers. HFA shadowers should not be able to manipulate their speech in an effort to negotiate social distance or expressing group membership. More detail on predictions for HFA shadowers is given in Phonetic Imitation Social factors may not be the only motivation for convergence. Exemplar models and other mechanistic theories such as Pickering & Garrod s (2004) interactive alignment model invoke automatic perception and production mechanisms to explain imitation. Under this mechanistic view, imitation occurs automatically and unconsciously as part of the evolution of our mental representations of language. To the extent that autism does not affect these mental representations, this mechanistic view should predict that the imitation by HFA shadowers will closely resemble that by TD shadowers. Exemplar models are one tool which researchers have used to examine phonetic imitation. Goldinger (1998) is often cited as one of the foundational studies in exemplar theory (e.g. Bybee, 2010; Johnson, 2006; Pierrehumbert, 2001). He provided evidence from word 3

14 shadowing experiments for detailed phonetic representations where the fine phonetic details of each experienced production are stored in memory. This conclusion was based on the finding that shadowers tend to imitate the pronunciation of auditory targets in word repetition. More specifically, shadowers imitated low frequency words more than high frequency words, and imitation increased with in-lab exposure to the target pronunciation. These effects were stronger in immediate shadowing than when shadowing was delayed by three or four seconds. Goldinger interpreted these results as evidence for a particular episodic or exemplar model of speech processing, based on previous work by Hintzman (1986). In this model, each encounter with a production of a word leaves a trace in memory. When a word is recognized, it activates neighboring phonetically similar traces, which may then influence subsequent pronunciations through a basic priming mechanism. Low frequency words are imitated more than high frequency words because they have fewer memory traces, allowing the exposure to a new production to more dramatically shift the aggregate representation for that word. High frequency words already have a large number of traces, so the comparably small number of exposures to a word in an experiment will have little effect on subsequent productions. Imitation of all words increases with recent exposure because more and more traces of the target are active in memory, and thus can exert greater influence on production. Goldinger claimed that his delayed repetition results confirmed the lexical mechanistic nature of imitation in his study. Shadowers who had to wait a few seconds before repeating the words did not imitate as much as shadowers who repeated right away. Any general tendency for imitation, social or otherwise, should have held despite the short lag in response time. Pickering & Garrod (2004) also proposed a mechanistic explanation for why speakers converge on multiple levels (e.g. phonetic, phonological, lexical, syntactic) during dialogue, but without specific reference to exemplar models. Their interactive alignment account is a broad psycholinguistic model of language processing during conversational 4

15 interaction, and an attempt to address the problem of basing language processing models on laboratory studies of monologue. The data which support their model are precisely the same kind of data that have been used in support of CAT, namely that conversation partners tend to converge on multiple levels of linguistic structure and behavior. Because they make reference to general priming mechanisms and mirror neurons rather than exemplars, they are not able to make specific predictions for word frequency or familiarity effects. The mechanistic and CAT approaches may seem incompatible, but some researchers have attempted to integrate them and determine the extent to which imitation is mechanistic versus mediated by social factors (Babel, 2009b; Pardo, 2006). Indeed, they appear to be two sides of the same coin. Mechanistic models predict a general tendency for speakers to imitate one another and attempt to provide a cognitive mechanism for how that imitation can take place. Speakers can be found to imitate even isolated words in a lab setting with no interlocutor present, which indicates a certain level of automaticity in imitation, all else being equal. In the real world, however, all else is rarely equal, and CAT provides an explanation for why imitation may be increased in certain situations or inhibited in others. For the purposes of the TD section of this study, we are focusing on the mechanistic side of imitation, specifically with reference to exemplar theory. That is, we are interested in using variation such as that between regional dialects to examine the architecture of mental representations of language. This approach is partially due to the methodology involved, that of a word shadowing task in a laboratory setting rather than observations of conversation between two interacting speakers. Even in a relatively non-interactive task, however, social factors cannot be ignored. Dialect variation is a social phenomenon as well as a linguistic one, and even in the laboratory setting participants may be making social judgments about the stimuli they hear. As such, we will be keeping both social and mechanistic explanations in mind, especially when we look at the results for the HFA shadowers. 5

16 1.2 Imitation and Dialect Frequency and Regional Dialect Goldinger s (1998) finding that word frequency affects imitation has been an important basis for exemplar models. According to the theory, exemplars which are encountered frequently and recently have high resting activation levels and will generally dominate production. Although several other variations on this basic setup exist (e.g. Johnson, 2006; Bybee, 2010), we will work with the mechanism proposed by Pierrehumbert (2001) because it is relatively explicit in specifying details of the model and its predictions for frequency effects. According to Pierrehumbert (2001), only exemplars within a fixed distance from the input are considered in recognition. This condition on activation means that when there is variation in the way a word is pronounced, a particular variant may only activate a subset of the preexisting exemplars for that word in memory. 1 Encountering a particular variant of a word may therefore strengthen the activation of the exemplars similar to that variant for future productions of the same word. In production, Pierrehumbert proposes that a label, such as a phoneme or word, is activated first, followed by a random sampling from the corresponding exemplar cloud which determines a target for production. This approach allows for abstract subparts of words such as phonemes or syllables to be integrated with the exemplar approach, thus capturing both the ability of speakers to use phonological units to create neologisms, for example, and the observable variation in the exact phonetic realization of such abstract units. While fine phonetic detail is retained in memory, the abstract category labels organize this detail into exemplar distributions for a given word, phoneme, or other unit. 1 This claim is dependent upon the parameters of the exemplar model specifying the distance from the input within which exemplar activation occurs. If this fixed distance parameter is too large, the distribution of all exemplars for the word may be activated, in which case local effects of activation of a particular exemplar, frequent or infrequent, will not occur. 6

17 The exemplars surrounding the production target are also activated and contribute to the production in accordance with their activation level. The result is entrenchment, through averaging to the mean of all activated exemplars around the target. High frequency words are imitated less in a word shadowing task such as Goldinger s (1998) because of this entrenchment process. In Goldinger s model, entrenchment occurs because of an echo in perception whereby any stimulus also activates neighboring exemplars. Whether you locate entrenchment in production or in perception, the basic result is the reduction of variance and reversion towards the mean of the distribution over time, due to mathematical averaging. To illustrate how entrenchment works, we can imagine a production target randomly selected near the edge of a given distribution. The neighbors of that target are also activated but there will be more neighbors with higher resting activations on the side of that target closest to the mean than on the side near the tail of the distribution. When the average of those neighbors is taken to determine the phonetic characteristics of the end production, it will tend to move closer to the mean than the original random target. Low frequency words have fewer, less activated exemplars in memory than high frequency words, which means that recent exposure to a particular exemplar can have a greater effect on subsequent productions. They may also exhibit less entrenchment than high frequency words, since entrenchment increases with repeated exposure and use. Word frequency, however, should not be the only place where we see the effects of entrenchment. In theory, the frequency of any variant within a category should have a similar effect on imitation. In the case of regional variation, dialects which are highly familiar to a particular speaker should behave similarly to high frequency words. A high frequency variant, when encountered, should activate a cloud of exemplars that are also frequently encountered and highly active, resulting in little change in production based on the input. A low frequency variant, on the other hand, may exert more influence on subsequent productions. 7

18 One key difference between words and dialect variants is that, at least in the version of exemplar theory put forward by Pierrehumbert (2001), words constitute linguistic categories with their own labels and exemplar distributions. Production starts with the selection of a word to be uttered, and a specific production target is selected through a random sampling over the exemplar distribution for that word. While word frequency effects such as Goldinger s deal with differences between the distributions of different words, dialect variation allows us to look within the distribution of exemplars for a single word. Dialect variation is useful for examining the behavior of exemplar distributions because we can talk about frequently encountered versus infrequently encountered dialect variants for a particular speaker. It is in this sense that I am using the terms familiar and unfamiliar, with regard to dialect. For our purposes here, a familiar dialect is one which the speaker has used and been exposed to frequently, defined based on residential history. A familiar dialect need not be the speaker s native dialect, but it should be a dialect with which the speaker has had substantial experience. I use the term unfamiliar not to imply a complete lack of experience with the dialect, but to indicate that a dialect is encountered less frequently relative to a speaker s familiar dialect. Familiar dialect variants will be likely to form the center of a speaker s distribution for any given word, while unfamiliar dialect variants will be at the periphery of the distribution. When a shadower hears a target from a familiar dialect in the shadowing task, it will activate exemplars near the mean of that shadower s distribution, changing the overall distribution only minimally. A less familiar dialect target should activate exemplars farther from the mean, slightly changing the overall distribution of exemplars for that word. The perception of a target from an unfamiliar dialect should thus exert more influence on subsequent productions. This hypothesis was tested in the shadowing experiment described below. Unfamiliar dialect targets may also be imitated more simply because there is more distance between them and the mean of the shadower s distribution. In other words, the 8

19 familiar variant of a word is also how that shadower would normally produce the word anyway, while there is more room for imitation of an unfamiliar dialect which is outside of their normal productions. This experimental design does not allow us to determine whether unfamiliar dialects are imitated more because of their place in the exemplar distribution or because they are farther away from the shadower s baseline productions. The two possibilities make the same predictions for familiar versus unfamiliar dialect imitation, and the practical distinction between them becomes even less clear when we consider the inherent variability in speech production, even at the center of the distribution. Even the baseline productions of a shadower exhibit some degree of variation, and the imitation effects in question here are small enough that imitation of a target near the mean should still be possible. Several studies provide evidence that less familiar dialects may be imitated more than highly familiar dialects, including those by Goldinger (1996, 2000) concerning familiar versus unfamiliar voices. He predicted that an odd or unfamiliar voice should activate fewer exemplars than a familiar voice, precisely because of its low frequency. It follows that subsequent productions may be more influenced by recent encounters with the unfamiliar voice than a familiar voice, since the familiar voice has built up a store of exemplars which can all have an influence on production, pushing productions toward the mean of the exemplar distribution. The findings reported by Nye & Fowler (2003) provide further evidence of the effect of linguistic experience on imitation. They asked shadowers to repeat stimulus recordings that approximated the phonotactic structure of English to varying degrees. Shadowers imitated stimulus recordings less when they approximated the phonotactic structure of English. Nye & Fowler interpreted their results as evidence that linguistic experience guides production. When shadowers could use their existing stores of exemplars of English words and syllables, their productions were dominated by previous experience. When the stimulus recordings did not resemble English words, the shadowers could not draw upon as 9

20 much of their stored experience, and were more influenced by the speech that they were shadowing. The current study extends the findings of Nye & Fowler to dialect variation. Just as linguistic experience with English words and syllables inhibited imitation, so should experience with a particular dialect. When shadowing a word of a less familiar dialect, shadowers have less experience to draw upon in production and may thus be influenced more by the target than when shadowing words in their highly familiar native dialect. At the same time, evidence suggests that some degree of familiarity with the target is necessary for imitation to occur. Pierrehumbert (2002) predicted that some exposure to a dialect is necessary for imitation, and that the most salient and frequent properties of a dialect should be the most likely to be imitated by speakers of a different dialect. This prediction arises out of the architecture of her model of perception and production, which requires a certain degree of previous exposure to a variety in order to accurately reflect its patterns of variation. In cases of limited exposure to a dialect, the most frequent and most salient features of that dialect should be the first features to build up the required distribution of exemplars in memory, with less frequent and less salient features requiring more exposure. Babel (2009a) made a similar claim that shadowers stay within their pre-existing production inventory while imitating. That is, if there are no exemplars already populating the space around the target, it should be impossible to produce a close imitation. Given Babel s claim, we would expect that highly unfamiliar dialects would not elicit imitation at all. This hypothesis cannot be tested here, given that the two dialects in question, the Midland and Northern dialects of American English, are relatively similar and geographically adjacent. While one dialect may be more familiar to certain speakers than the other, it is unlikely that any speaker in this study would be completely unfamiliar with the other dialect. For the purposes of this study, even the unfamiliar dialect must be familiar enough to be represented in the distribution of exemplars for a given speaker, but not so familiar 10

21 that it occupies the center of that distribution Dialect Processing In addition to the predictions that follow from exemplar theory, previous research also provides some information on the processing of dialect variation. There is a processing benefit for familiar dialects. Unfamiliar regional dialects can create a temporary lag in response time in a lexical decision task (Floccia et al., 2006), and unfamiliar dialects are less intelligible in noise than familiar dialects (Clopper & Bradlow, 2008; Adank et al., 2009). Sumner & Samuel (2009) also found that mere exposure to a dialect can improve performance on lexical decision tasks. In their study, they found that r-ful listeners in New York were faster and more accurate in making a lexical decision in responses to r- less stimuli if the listeners had a large amount of previous exposure to r-less speech. This result shows that even when a speaker does not normally produce a variant, that variant may still be familiar enough to facilitate processing. Other studies, such as Niedzielski s (1999) work, show that social information may also impact how speakers perceive regional dialects. Merely changing the label associated with a speaker from Michigan to Canadian changed the vowel quality that was perceived by listeners. Her findings highlight the necessity of paying attention to social categories when dealing with variation such as that between regional dialects. Niedzielski (1999) also reported that speakers from Detroit apparently did not perceive some dialect features of the Northern Cities Shift in their own productions. This result shows that the perception of dialect variation is not always precisely in line with acoustic measurements of variation. Sociolinguistic category labels can shift the perception of speech, and speakers are not always aware of dialect features that can be measured acoustically. There is some evidence that speakers may not be overtly aware of the differences between the Northern and Midland dialects. Northern speakers may have a high degree of linguistic security, meaning that they consider their own speech to be close to the standard 11

22 (Preston, 1989). Moreover, Clopper & Pisoni (2004) reported that naïve listeners have difficulty distinguishing between the Northern, Midland, and Western dialect regions. These findings suggest that at least some speakers may not be overtly aware of differences between the Northern dialect and the Midland dialect, in the sense that they are not able to categorize them into distinct groups. This is not to say, however, that listeners could not distinguish in some ways between Northern and Midland tokens given a different task. It may well be that Northern and Midland tokens could elicit different responses from listeners on rating tasks, such as determining how friendly or educated a talker sounds, for example. Even if speakers may not be aware of them, production studies have revealed differences between the Northern and Midland dialects. The most famous of these may be the Northern Cities vowel shift, found in the urban areas of the Northern dialect region (Labov et al., 1973; Wolfram & Schilling-Estes, 1998). Figure 1.1, adapted from Clopper & Pisoni (2004) and Wolfram & Schilling-Estes (1998), shows a schematic of the vowel movements involved in the Northern Cities Shift. The low vowels /æ/, /A/, and /O/ are moving forward in a chain shift, while the vowels /I/, /E/, and /2/ move back. Figure 1.1: Vowel plot of the Northern Cities Shift. The Midland dialect, on the other hand, is characterized by a merger of the /A/ and /O/ low-back vowels (Wolfram & Schilling-Estes, 1998; Labov et al., 2006). It is also char- 12

23 acterized by a fronted /u/ and /o/, although /u/ is also moving forward in the North (Preston, 2003). While Pierrehumbert s prediction about dialect familiarity was based on exemplar theory, Trudgill (1981) makes a similar prediction based on social saliency. He claimed that socially salient variables should be imitated more than less socially salient variables. To the extent that perceptually salient and frequent variables are also likely to become socially salient, these two predictions go hand in hand. Given the documented differences between the Midland and Northern dialects, we might expect some of the shifted vowels, such as /E/, /æ/ and /A/ to exhibit more imitation by Midlanders shadowing Northern targets, while /u/, /o/, and the merged /A/-/O/ might elicit more imitation by Northerners shadowing Midland targets. Of course, not all of the documented differences between the two dialects can be assumed to be socially salient, especially given the results of Clopper & Pisoni (2004) showing that listeners tend to group the two dialects together. Babel (2009a) found that the vowel /æ/ was imitated more than other vowels, and Pardo (2010) found the same for both /æ/ and /A/, so we should also expect high degrees of imitation for these low vowels, regardless of dialect. 1.3 Regional Dialect Imitation and Autism Imitation and Autism While much work has been done looking at spontaneous phonetic imitation in TD populations, less is known about autism and the imitation of phonetic detail. Action imitation is impaired in children with autism (Williams et al., 2004) and in adolescents and young adults with HFA (Stieglitz Ham et al., 2008). Children with Ausitsm Spectrum Disorder (ASD) have difficulty with both motor and vocal imitation (Stone et al., 1990). Given that ASD is also generally characterized by language deficiencies (American Psychiatric Association, 1994), we might expect imitation of speech to be impaired in individuals with 13

24 HFA. Some studies, however, show that at least some imitation of phonetic detail must be possible for individuals with HFA. Volden et al. (2007) report that high-functioning children and adolescents with ASD were able to vary their language register and style for different interlocutors (e.g. foreigner talk or baby talk), but that they were not able to do so at the same level as TD controls. In another study, however Volden & Sorenson (2009) showed that this same population was able to distinguish between nice and bossy requests just as well as TD children. Adjusting speech to express different registers or to talk to different listeners requires both a control of phonetic detail and a certain level of social awareness in linguistic interactions. Furthermore, children with ASD adopt the phonetic details of the dialect in their surroundings, although their acquisition differs from TD children. While TD adolescents tend to adopt the dialect features of their peers (Chambers, 2002), children with ASD adopt speech more like their mothers, while their TD siblings sound more like their peers (Baron- Cohen & Staunton, 1994). Language acquisition in general and the acquisition of dialect features requires phonetic imitation to take place. We also have some evidence that individuals with autism have normal or even enhanced attention to phonetic detail. There is a high incidence of absolute pitch in autism (Heaton et al., 1998) and an enhanced discrimination of pure tones (Bonnel et al., 2003). More general auditory sensitivity seems to be related to Autism Spectrum Quotient (AQ) scores (Mottron et al., 2006; Happé & Frith, 2006), but it is not necessarily associated with an autism diagnosis (Rogers & Ozonoff, 2005). Individuals with ASD can also recognize voices and link them to specific talkers (Boucher et al., 2000), a task which requires attention to fine details of phonetic variation. Further complicating matters is the phenomenon of echolalia, or the verbatim repetition of words or phrases uttered by another speaker, with similar intonation (Tager- Flusberg et al., 2005). Echolalia is one of the classic symptoms of autism in children (Kan- 14

25 ner, 1946). In many cases these echoes can be very close acoustic matches to the model utterance, specifically in terms of intonation (Local & Wootton, 1995). At the same time, other studies show that autistic individuals have difficulty perceiving the emotional intent of speech (Golan et al., 2007; Imaizumi et al., 2009). Hubbard & Trauner (2007) report that participants with ASD had difficulty producing different emotional colors in speech. Similar difficulties have been reported for the perception and production of prosody (Paul et al., 2005; McCann et al., 2007). Even these results, however, are mixed, as Järvinin-Pasley et al. (2008a,b) reported that people with ASD can recognize and judge specific prosodic contours just as well as TD peers. It appears that while individuals with ASD may have some difficulty interpreting prosody and emotion in speech, they are not impaired in all circumstances. Given these mixed results, we should expect that imitation by HFA shadowers will be either identical to that by TD shadowers or only slightly impaired. Previous results showing enhanced pitch perception and imitation in intonation in echolalia suggest that HFA shadowers may produce more exact repetitions of f0 than TD shadowers Dialect and Autism Evidence from a recent study of dialect and HFA by Clopper et al. (2012) gives reason to expect that TD and HFA shadowers may produce different patterns of imitation when shadowing talkers from different dialects. According to their results, listeners with HFA can perceive indexical cues to dialect and classify talkers by dialect. This ability should not be surprising given the ability of HFA listeners to perceive and manipulate fine-grained phonetic detail. Clopper et al. (2012) also administered a language attitudes task based on classic studies such as Giles (1970) and Luhman (1990) in which listeners rate talkers on a range of personal attributes such as intelligence, level of education, etc. TD listeners made more distinctions between talkers of different regional dialects than HFA listeners, 15

26 indicating that HFA listeners had difficulty in associating stereotyped social information with dialect categories. These results fit with the Weak Central Coherence theory which proposes that autism is characterized by a processing style with a weakened ability to integrate different sources of information but retain the ability to focus on detail (Happé & Frith, 2006). HFA listeners can hear the differences between talkers of different dialect groups, but they have difficulty combining that information with socially defined language attitudes. There is also evidence that people with autism have an impaired ability to integrate other kinds of linguistic information. Participants with ASD employ semantic context less when resolving orthographic and auditory ambiguities (Happé, 1997; Jolliffe & Baron-Cohen, 1999) and exhibit a reduced ability to use lexical information in speech perception (Stewart & Ota, 2008). The influence of sociolinguistic preconceptions about a talker s dialect should therefore have less influence on imitation of a dialect by shadowers with HFA than for typically developing shadowers. We should be able to see this difference between the two shadower groups in the degree of imitation of Midland versus Northern talkers. The TD listeners from Clopper et al. (2012) rated Midland talkers highest of all dialect groups on all rating scales (friendliness, reliability, intelligence, successfulness). HFA listeners also rated Midland talkers highest in intelligence but did not differentiate between dialect groups on the other three rating scales. CAT predicts that the TD shadowers should imitate the highlyrated Midland talkers more than Northern talkers. We should expect that such an effect would be lessened for HFA shadowers. Based on these findings, we might expect that the HFA shadowers may exhibit generally impaired imitation, although such an impairment should reflect social deficits rather than auditory or perceptual problems. The HFA shadowers might be particularly good at imitating f0 as compared to other aspects of the speech signal. To the extent that social information affects imitation in this shadowing task, those effects should be less apparent 16

27 for HFA shadowers than TD shadowers. 17

28 Chapter 2 Imitation in Typically Developing Adults 2.1 Experiment 1: TD Word Shadowing Methods Participants 39 undergraduate linguistics students at The Ohio State University received partial course credit for their participation in Experiment 1. Eight participants were excluded for various reasons, including a history of speech or hearing problems (N = 2), a native language other than or in addition to English (N = 5), or because of a residential history outside of the three dialect groups needed for the study (N = 1). The remaining 31 TD participants were grouped according to residential history into three dialect regions: Midland, Northern, and Mobile. Northern and Midland participants had lived exclusively in their respective dialect region before beginning their studies at Ohio State. Mobile participants were those who had lived in more than one dialect region before arriving at the university, where one of those regions was the Midland or the North. 1 Table 2.1 shows the breakdown of the three shadowing groups by gender. Figure 2.1 shows the locations of the Northern and Midland dialects of American English. The star indicates the location of Columbus, OH, where the study was conducted. 1 While geographic criteria are certainly a convenient way of classifying speakers into dialects, they are by no means the only way of characterizing regional variation. Dialect boundaries, however, are gradient both geographically and acoustically, and the issue of gray areas in classification could arise with any method. In this case, we felt that classifying participants phonetically would run the risk of creating groupings based on experimenter bias and preconceptions of dialect features. Geography provided a comparatively principled criterion for classification. 18

29 Table 2.1: Participant groups in Experiment 1. dialect group number of shadowers example residential history Midland 15 [9m/6f] e.g. Columbus, OH Northern 8 [3m/5f] e.g. Cleveland, OH Mobile 8 [2m/6f] e.g. Midland and Southern regions As all students were recruited in Columbus, Ohio, they were residing in the Midland region at the time of this study. All participants were between the ages of 18 and 25 years old. Figure 2.1: Map of the Northern and Midland dialect regions. Of the eight Mobile TD shadowers, five had lived in both the Midland and the Northern regions, two in the North plus another region, and one in the Midland plus another region. Table 2.2 shows the details of the residential history for the Mobile shadowers. Table 2.2: Residential history for Mobile TD shadowers. dialect groups number of shadowers Midland, North 4 Midland, North, South 1 Midland, International 1 North, International 1 North, South 1 19

30 Materials The experimental materials consisted of 55 monosyllabic CVC English words taken from the Indiana Speech Project corpus (Clopper et al., 2002). Each word was produced by three female Midland target talkers and three female Northern target talkers, for a total of 330 target productions. The word list was selected to sample the American English vowel space, and included five words for each of 11 vowels: /i/, /I/, /E/, /e/, /æ/, /u/, /o/, /U/, /O/, /A/, and /2/. All words were previously rated as having an average familiarity of 6.0 or greater on a 7-point scale by undergraduate students (Nusbaum et al., 1984). Care was taken to avoid words where a dialect difference in vowel quality could result in lexical ambiguity. The complete word list is provided in Appendix A. Figure 2.2 shows the mean formant values for the stimulus words used in Experiment 1. Formant values were measured at vowel midpoint and z-score normalized. The Northern targets exhibit some of the characteristics of the Northern Cities Shift. The Northern /æ/ is raised, and the Northern /I/, /E/ and /2/ are lower and farther back than their Midland counterparts. In fact, the Northern /æ/ is almost in the same location as the Midland /E/. Both the Midland and Northern targets have a fronted /u/, while only the Midland targets have a fronted /o/ as well. The Midland /A/ and /O/ are also slightly closer together than their Northern counterparts, as would be expected given a low-back merger. The Northern high vowels are all higher than the Midland high vowels, although it is not clear why this should be the case. Task Experiment 1 was based on the word shadowing task in Goldinger (1998). Shadowers first read all 55 target words aloud, without exposure to any auditory materials. These recordings served as a baseline for the productions of each shadower and will be referred to as the baseline productions throughout this paper. Each shadower then heard all

31 Figure 2.2: Vowel plot for Midland and Northern target vowels. Mean Target Formant Values first formant measured at vowel midpoint (z-score norm.) i i e e Midland North ɪ ɪ æ ɛ ɛ u u æ ʊ ʊ a ʌ ʌ a o o ɔ second formant measured at vowel midpoint (z-score norm.) target productions one at a time and repeated each target aloud as quickly and naturally as possible. The targets and repetitions will be referred to as target and shadowed productions, respectively. Shadowers were not explicitly told to imitate the target productions. The instructions given to the shadowers were to repeat each word as quickly and naturally as possible. Targets were presented in a random order, with no blocking by target talker or by target talker dialect. Shadower responses were recorded using a head-mounted Sennheiser HMD 280 microphone and a Marantz PMD671 solid state recorder in a sound-attenuated booth. The microphone signal was amplified by a tube microphone preamplifier (Applied Research and Technology). Target stimuli and shadowed productions were recorded simultaneously as separate channels of the same 44.1 khz, 16 bit stereo audio file. 21

32 2.1.2 Measures The acoustic measures listed in (1) were obtained for each shadowed production, baseline production, and target production. (1) Acoustic measures a. Vowel quality, i.e. first (F1) and second (F2) formant values, at vowel midpoint b. Duration of onset consonant, vowel, and coda consonant c. Fundamental frequency (f0) at vowel midpoint. d. F0 trajectory or change in f0 between points at 1/4 and 3/4 of the total vowel duration. 2 The response time for each trial was also calculated, measured from the onset of the target production to the onset of the shadowed production. Response time was calculated by hand from the stereo recording. An acoustic imitation score was calculated separately for each of the acoustic measures. For all of the measures, acoustic imitation on each trial was defined as the distance from the shadowed value to the target value, subtracted from the distance from the baseline value to the target value. The formula for calculating acoustic imitation is given in (2). (2) acoustic imitation = X targ - X base X targ - X shad The acoustic imitation score shows how much closer the shadowed production is to the target production as compared to the baseline for that shadower. In (2), a negative acoustic imitation score reflects divergence away from the target, and a positive acoustic imitation score reflects convergence to the target. This measure is equivalent to Babel s (2009a) Difference in Distance, although her method reverses the values so that imitation is indicated by negative numbers. 2 A positive number for f0 trajectory indicates a rising f0, negative indicates falling f0. 22