Inter-speaker articulatory variability during vowel-consonantvowel sequences in twins and unrelated speakers Melanie Weiricha) Zentrum f€ ur Allgemeine Sprachwissenschaft, Sch€ utzenstrasse 18, 10117 Berlin, Germany

Leonardo Lancia Max Planck Institute for Evolutionary Anthropology, Deutscher Platz 6, 04103 Leipzig, Germany

Jana Brunner Universit€ at Potsdam, Department Linguistik, Karl-Liebknecht-Strasse 24-25, 14476 Potsdam, Germany

(Received 31 August 2011; revised 5 September 2013; accepted 9 September 2013) The purpose of this study is to examine and compare the amount of inter-speaker variability in the articulation of monozygotic twin pairs (MZ), dizygotic twin pairs (DZ), and pairs of unrelated twins with the goal of examining in greater depth the influence of physiology on articulation. Physiological parameters are assumed to be very similar in MZ twin pairs in contrast to DZ twin pairs or unrelated speakers, and it is hypothesized that the speaker specific shape of articulatory looping trajectories of the tongue is at least partly dependent on biomechanical properties and the speaker’s individual physiology. By means of electromagnetic articulography (EMA), inter-speaker variability in the looping trajectories of the tongue back during /VCV/ sequences is analyzed. Results reveal similar looping patterns within MZ twin pairs but in DZ pairs differences in the shape of the loop, the direction of the upward and downward movement, and the amount of C 2013 Acoustical Society of America. horizontal sliding movement at the palate are found. V [http://dx.doi.org/10.1121/1.4822480] PACS number(s): 43.70.Bk, 43.70.Aj [CHS]

I. INTRODUCTION

Since the search for acoustic or articulatory invariant patterns in speech started, inter-speaker variability has been an important issue in understanding not only the speech production process but also speech perception. Ladefoged and Broadbent (1957, p. 98) state that “the idiosyncratic features of a person’s speech” may “be a part of an individual’s learned speech behavior” or might be “due to anatomical and physiological considerations.” These two influencing parameters can be described and specified as biological determinants (i.e., genetics, physiology, biomechanics) and non-biological determinants (i.e., social environment, learning, linguistic factors). The aim of this study is to explore how articulatory inter-speaker variability is influenced by biological determinants and more precisely by physiological parameters such as vocal tract morphology, tongue shape, and muscle characteristics. To do so, the speech of monozygotic (MZ) and dizygotic (DZ) twins will be analyzed because, as will be shown in detail here, the reasons for speaker specific variability can be differentiated in terms of physiologically determined factors and learned articulatory behavior. The analysis will focus on a particular kind of articulatory movement, the looping movement of the tongue back in velar stop production. The reason for investigating this kind of movement is that it has been argued to be dependent on the speaker’s physiology and also

a)

Author to whom correspondence should be addressed. Present address: Friedrich-Schiller-Universit€at Jena, Institut f€ ur Germanistische Sprachwissenschaft. Electronic mail: [email protected]

3766

J. Acoust. Soc. Am. 134 (5), November 2013

Pages: 3766–3780

to be controlled by the speaker, so that it could also be considered learned behavior. A. Twins’ speech

Twin studies are a common type of investigation in the field of psychology. The origins of twin studies go back to the late 19th century and are ascribed to Sir Francis Galton (Galton, 1876), even though it is not clear whether he was aware of the difference between monozygotic twins (who are genetically identical) and dizygotic twins (who share around 50% of their genes on average). Nowadays the systematic comparison of the within-pair similarity of MZ twins with that of DZ twins is a standard procedure in the field of behavioral genetic research. The aim is to investigate individual differences and to explain the variation in terms of two possible influencing factors: (1) Genes and physiology and (2) environmental parameters. The latter factor refers to social-environmental factors that contribute to the resemblance between individuals who grow up in the same family. The Equal Environments Assumption (EEA) assumes that MZ and DZ twins share the same amount of environmentally based similarity. This crucial assumption has been investigated intensively, and mislabeled twin pairs (i.e., DZ twins that grew up as MZ twins and MZ twins that grew up as DZ twins) have shown the validity of this assumption (Scarr and Carter-Saltzman, 1979). Additionally, the study by Koeppen-Schomerus et al. (2003) regarding language and cognitive measures of 2- and 3-yr-old twins and non-twin siblings shows that the estimated amounts of shared environment are more than twice as large for twins (DZ and MZ) as compared to non-twin siblings. The only difference in terms

0001-4966/2013/134(5)/3766/15/$30.00

C 2013 Acoustical Society of America V

of the two factors genes/physiology and environmental parameters between MZ and DZ twins who are still living together and share their social environment is the difference in their genetic or physiological similarity. Thus the assumption can be made that if MZ twins are more similar in an investigated parameter than DZ twins, it points to the importance of the genetic influence (if environmental parameters are kept constant). Several studies have shown that anatomical and physiological characteristics are more similar in MZ twin pairs (also called identical twins) than in DZ twin pairs (also called non-identical twins) (e.g., Lundstr€om, 1948; Kabban et al., 2001). Lundstr€om’s early medical dissertation on anatomical variation in twins showed that identical twin pairs reveal less variation in the size and position of the jaw and the teeth than non-identical twin pairs. Langer et al. (1999) found in their ultrasound study including 28 monozygotic twin pairs and 13 dizygotic twin pairs no MZ pair with strikingly discordant thyroid volumes but one such pair among DZ twins. In addition, tooth size and occlusal morphology has been shown to be remarkably similar in monozygotic twins, while differences were found in dizygotic twin pairs and even more variability was apparent in unrelated speakers (Kabban et al., 2001). Eguchi et al. (2004) also found a high genetic contribution to the variation in dental arch width, length, and palatal height in their study on 78 male and female Australian MZ and DZ twin pairs. While in the field of psychology twin studies are very common, comparing the within-pair variability of DZ and MZ twins in speech is rather new. Within the debate about the innateness of language and linguistic knowledge (e.g., Chomsky, 1975), twin studies have been conducted in the fields of speech acquisition and speech pathology (Locke and Mather, 1989; Simberg et al., 2009; Ooki, 2005). However, speaker-specific characteristics of normal speech are a less frequent research topic (see Loakes, 2006, p. 41). Acoustic and perceptual studies on twins’ speech have shown that MZ twins are more similar than same-sex DZ twins or age-matched siblings in their acoustic output: E.g., mean F0 (Przybyla et al., 1992; Debruyne et al., 2002), voice quality parameters (van Lierde et al., 2005) and coarticulatory patterns (Nolan and Oh, 1996; Whiteside and Rixon, 2003). While it seems to be very difficult for unfamiliar listeners to distinguish MZ and DZ twins by listening to just one word (Weirich and Lancia, 2011), even MZ twins can be distinguished in perception tests with familiar listeners (Whiteside and Rixon, 2000) and in acoustic analysis on formant patterns (Loakes, 2006) or by using automatic speaker recognition systems (K€unzel, 2010). Recently, a comprehensive study on acoustic, perceptual but also articulatory parameters in the speech of German MZ and DZ twin pairs was conducted (Weirich, 2012). The subject group was chosen such that both twin types show the same amount of shared environment but differ with respect to their physiological similarity. This was tested through several measurements (e.g., DNA test, investigation of tongue size and palate contour, questionnaire on childhood and attitude toward being a twin; see Weirich, 2012 for details). The J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

study did not find differences in the amount of within-twin pair variability between DZ and MZ twins in the acoustics and articulation of the stressed corner vowels /a, i, u/ (target tongue positions, stable formant patterns in the middle of a vowel) (Weirich, 2010, 2012). In addition, a perception test revealed that neither DZ nor MZ twin pairs could be distinguished by listeners unfamiliar with their voices, although the listeners could easily distinguish sex and aged-matched unrelated speakers (Weirich and Lancia, 2011). However, MZ and DZ pairs differed in their within-twin pair variability with regard toÐ the articulatory realization of the phoneme contrast /s/ vs / / (Weirich and Fuchs, 2011). While MZ twins reveal very similar articulatory realizations of the contrast, DZ twins differ in their amount of horizontal and vertical distance between the articulatory target positions of the two sibilants. Results were interpreted in terms of a strong influence of the palatal shape, in particular the steepness of the alveo-palatal ridge, on the articulatory realization of this contrast. Furthermore, Weirich (2012) showed that while MZ twins are very similar in dynamic formant patterns (i.e., F2 and F3 transitions) in sibilant–schwa sequences, DZ twins differ from each other. In contrast, the investigated static parameters in sibilants (spectral center of gravity and main spectral peak) and schwa realizations (formants measured in the stable part of the vowel) do differ in both MZ and DZ twins. One issue discussed within this framework was the question of whether a difference exists in the influence of physiological parameters on inter-speaker variability between static and dynamic parameters: Within-pair variability in static parameters turned out to be similar in DZ and MZ pairs, but in dynamic parameters, was larger in DZ than in MZ pairs. Thus dynamic parameters might be influenced more by physiological properties. In addition, other studies have found that MZ twins were remarkably similar in dynamic coarticulatory patterns (Whiteside and Rixon, 2003; Nolan and Oh, 1996). Therefore the present investigation looks at dynamic patterns in speech; however, this time the focus lies on articulatory movements. By investigating articulatory movements in twins and unrelated speakers, we can (1) analyze the influence of physiology on articulatory variability and (2) take the dynamic nature of speech into account. Thus the aim of the present study is to investigate to what degree articulatory patterns are learned by a shared environment and to what degree they are influenced due to a speaker’s physiology. Studying inter-speaker variability in unrelated speakers, MZ twins, and DZ twins allows us to make this distinction: If an articulatory pattern is similar within MZ twin pairs but differs within DZ twin pairs and unrelated speakers, it can be assumed to be due to the physiological properties of the vocal tract (which is very similar in MZ twins but differs in DZ twins and between unrelated speakers). B. Looping movements

Nolan et al. (2006) assume that the speech signal contains (a) linguistically determined targets, which are constrained by the shared language system, and (b) organically Weirich et al.: Articulatory variability in twins’ speech

3767

determined and speaker-specific transitions, which link the adjacent linguistic targets. For an investigation of the influence of physiology, the dynamic trajectory (rather than the static parameters) is therefore of primary importance. A specifically interesting kind of trajectory is the looping movement during velar stops. A loop is a curved movement of the tongue back in the production pattern of VCV sequences where C is a velar stop consonant. Loops are particularly obvious in a symmetric vowel context. These trajectories are remarkable because they are not simply produced on a straight path from the vowel to the consonant target and back along another straight line to the next vowel target; rather an elliptical movement of the tongue back—a loop—is found (Kent and Moll, 1972; Mooshammer et al., 1995; Hoole et al., 1998; L€ ofqvist and Gracco, 2002; Geng et al., 2003; Perrier et al., 2003; Brunner et al., 2011). Curved paths, in speech as well as in other human movements, have been shown to be potentially explained by anatomical factors and muscle mechanics (cf. Flanagan et al., 1993; Gribble and Ostry, 1996; Gribble et al., 1998 for arm movement; Perrier et al., 2003; Perrier and Fuchs, 2008 for orofacial movements). Some studies therefore argue that loops are the result of the biomechanical arrangement of the muscles (e.g., Perrier et al., 2003). In this case, they should be very similar in MZ twins but differ in DZ twins and unrelated speakers. Other studies have argued that the loops are actively controlled movements (L€ofqvist and Gracco, 2002; see following text for more details). In this case, they should present instances of learned behavior, and the degree of variability within a twin pair should only minimally differ for MZ vs DZ pairs. Houde (1967) and Kent and Moll (1972) suggest that the air pressure behind the constriction, which can be assumed to be controlled in stops, may have an influence on the shape and size of a loop. Coker (1976) and Houde (1967) assume that the looping movement might be used to sustain voicing through active cavity enlargement. Houde (1967) claims that “the direction of the movement during closure is consistent with an increase in oral pressure, and as in the case of labial closures, a compliant element is required in the oral cavity during the voiced palatal stop in order to sustain voicing. The passive reaction of the tongue may provide that required compliance” (p. 133). Ohala (1983) suggests that the looping pattern could serve as an active cavity enlargement and a compensation for other factors that disfavor voicing in velars. Mooshammer et al. (1995) analyzed lingual movement during VCV sequences with varying vowel contexts and manners of articulation of two speakers of German by means of electromagnetic articulography. They found that the articulatory loops were larger for the sequence with an unvoiced stop consonant (/aka/) than for the sequence with a voiced consonant (/aga/), thereby challenging Ohala’s explanation. To analyze the possible influence of aerodynamic forces further, Hoole et al. (1998) conducted a study comparing the production of velar consonants during normal and ingressive speech. In both cases, forward articulatory loops were found but of a smaller size in ingressive speech, suggesting that aerodynamics 3768

J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

does indeed seem to influence tongue movements. The question whether loops containing voiced stops differ from loops containing unvoiced stops in terms of their sensitivity to individual physiology is one aspect that will be addressed in the further analysis. If this is the case, the comparison between the degree of variability within MZ and DZ pairs should be affected and differ between /VkV/ and /VgV/ stimuli. L€ofqvist and Gracco (2002) also investigated tongue movements in V1CV2 sequences with a lingual stop consonant and the vowels /a, i, u/ in four subjects by means of a magnetometer system. However, they assume that physical factors such as aerodynamics and biomechanics only play a minor role in the trajectory shape and that the tongue movement during the closure is actively controlled. The authors interpreted their finding of a high velocity at the onset of the oral closure in terms of a virtual target that lies beyond the palate. The resulting tissue compression ensures the needed airtight closure for the production of a stop consonant. The authors suggest that the use of such a control strategy would guarantee a closure regardless of contextual variability. This closure is seen as the target for stop consonants, and because of the magnitude of the tongue movement found during closure, the authors suggest it to be actively controlled rather than solely due to biomechanical and/or aerodynamic factors. Perrier et al. (2003) explicitly investigated controlled and uncontrolled influences on the looping movement. They tried to determine if the exact size and shape of the loops are planned by the speakers or are affected by the anatomical and physiological arrangements of the tongue muscles. Using a tongue model, they conducted V1CV2 movement simulations, where C is a velar consonant and V is /a/, /i/ or /u/. Basically, the model used in this study consists of hard structures (like palate, teeth, and bones) and a soft body that represents the tongue with seven muscles that can be controlled. Vowels and consonants were specified in terms of targets. The resulting trajectories of the different VCV simulations were inspected. Results showed that if V1 is /a/ or /u/, the trajectory forms a forward loop that starts before the consonant closure and continues to slide along the palate during the closure, while the direction of the path is backward for sequences where V1 is /i/. Note that this finding is contrary to the results of Houde (1967), who found forward loops for V1 ¼ /i/. Thus a more variable pattern for V1 ¼ /i/ in the orientation and shape of the loops depending on speaker or language might be assumed. The amount of vertical movement during the consonantal closure turned out to be dependent on V1 and ranges from 5 mm for /a/ to 3 mm for /u/ to 2 mm for /i/. The authors assume an influence of the shape of the palate on the horizontal component of the curved trajectories. They suggest that one important point is the distance the tongue has to travel between the first palate contact and reaching the target position. This is especially relevant for the present study because the amount of horizontal movement might be constrained by the (individual) palatal contour (i.e., physiology). Based on their findings, Perrier et al. (2003) conclude that in addition to the influence of the neighboring vowel targets, the shape of the trajectory is also Weirich et al.: Articulatory variability in twins’ speech

affected by biomechanical properties of the tongue (the passive tongue elasticity, the muscle arrangements within the tongue, the force generation mechanism), and a general optimization principle that plans the entire trajectory, as proposed by L€ ofqvist and Gracco (2002), is not necessary to explain the shape of the trajectory. Perrier et al. (2003) emphasize that the presence and general shape patterns of the loops result, at least in part, from tongue biomechanics and muscular anatomy, and the upper portion of the loop and the amount of vertical sliding during the palatal contact might be influenced by speaker-specific palatal shapes. The question whether V2 in V1CV2 sequences has an effect on inter-speaker variability caused by physiological factors will be examined in the present analysis which includes /a/ and /u/ as V2. The study of Birkholz et al. (2011) also supports the important role of biomechanics as a cause of the loops. However, they found that loops also occur in V1-V2-V1 sequences without any consonant involved. From this, they conclude that muscular effects are most important for the resulting looping movements while air pressure as well as tongue-palate interactions only play a minor role. However, even though loops can also appear in sequences without a velar consonant and thus tongue-palate contact does not seem to be necessary, the tongue-palate contact that is apparent in the production of VCV sequences might be a potential impact factor with regard to inter-speaker variability in loops. This issue is analyzed and discussed in the present study. The preceding discussion shows that for the investigation of the influence of a speaker’s physiology on articulatory movements, loops are an ideal phenomenon because they seem to be influenced by both the physiology but also learned articulatory behavior. Here the specific shape of the looping trajectory is of special interest. Therefore the curvature of the movement (and not the horizontal and vertical position) will be compared between the speakers. Using curvature minimizes the interfering effect of sensor positioning on inter-speaker variability. We will come back to the advantages of curvature in the beginning of the method section. C. Hypotheses

The present study investigates the influence of physiology and biomechanics on articulatory behavior by comparing speakers with different degrees of anatomical similarity: (a) Monozygotic twin pairs, who have identical genes and shared physiology, (b) dizygotic (same sex) twin pairs, who share only 50% of their genes on average and thus show measurable differences in their physiological parameters, and (c) speaker pairs formed out of the unrelated twins matched in sex and age. In other words, the aim of the following analysis is to find out if speaker-specific biomechanical and physiological characteristics influence the articulatory looping movements, which may therefore be more similar in physiologically identical MZ twins than in DZ twins or unrelated speakers. Two assumptions are made: J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

(1) MZ twins are more similar in their physiology than DZ twins and unrelated speakers. (2) MZ twins and DZ twins share the environment so that they should show the same learned behavior. Unrelated speakers do not share the environment and can therefore differ in their learned behavior. Therefore it is hypothesized that if looping movements are influenced by a speaker’s physiology, MZ twins should be more similar than DZ twins and unrelated speakers in their articulatory movements of the tongue back during /VCV/ sequences. In addition, the potential impact of the second vowel (i.e., /a/ or /u/) and the presence of voicing in the stop (i.e., voiced or unvoiced) will be investigated. II. METHOD

We conducted acoustic and articulatory recordings of four /V1CV2/ sequences—where V1 is /a/, V2 is /a/ or /u/, and C is /g/ or /k/—by means of electromagnetic articulography (EMA) with ten German speakers. For the acoustic recordings, a Sennheiser Mkh 20 P48 microphone was used (48 kHz sampling rate). The present study focuses on the articulatory data, in particular, on the shape of the looping movement of the tongue back acquired over the whole sequence for each speaker. First, the specific shape of the looping trajectories of each speaker and each twin pair were examined visually. Second, curvature was calculated for each articulatory trajectory, following the method proposed by Tasko and Westbury (2004). Curvature is defined as the rate of change of a trajectory’s direction with respect to arc length (Hurley, 1980). For example, a circle has a constant curvature, and a circle with a small radius has a higher curvature than a circle with a bigger radius. The magnitude of curvature increases as a trajectory becomes more sharply curved and decreases as a movement becomes straighter. The sign of curvature signifies whether the curve is directed clockwise or counterclockwise. By analyzing the curvature data and not the positional data, the shape of the trajectory can be examined best. The comparisons are not influenced by potential differences in coil position or the general spatial position of the loops. This makes it possible to focus on the analysis on the shape of the trajectories. In a third step, the curvature trajectories were time aligned through the registration method proposed in the framework of functional data analysis (FDA). With this technique (Lucero et al., 1997), it is possible to compare the shapes of various trajectories regardless of their timing, which can be affected by many variables, such as speech rate, which would introduce differences across the trajectories that are not relevant for our study. Finally, mean absolute differences were measured between the aligned curvature trajectories of each pairwise speaker comparison and each /aCV/ sequence separately to conduct a statistical analysis. Speaker pairs were formed with data coming either from MZ twin pairs, from DZ twin pairs, or from unrelated speaker pairs. Finally, a linear mixed model (LMM) was calculated to look for a significant effect of physiology on the pattern of the looping trajectory. This is outlined in more detail is Sec. III D. Weirich et al.: Articulatory variability in twins’ speech

3769

A. Subjects and speech material 1. Physiological similarity of the subjects

Ten German native speakers (five twin pairs) from Berlin took part in the investigation: Three MZ pairs (two female, one male) and two DZ pairs (both female). The genetic similarity (zygosity) of these twin pairs was determined by a genetics laboratory through a genotypic comparison based upon 16 different genetic markers. Monozygotic twin pairs are 100% genetically identical. If a twin pair differs in any of the 16 DNA markers, they must be dizygotic. When a reasonable number of markers (here 16) reveals no differences, it can be concluded that the twin pair is monozygotic (Spinath, 2005). An overview of the characteristics of our subjects and the influence factor genetic identity (zygosity) is given in Table I (also including other information). A code for the different twin pairs was generated that gives information about the genetic identity (MZ or DZ), the sex (M or F) and a running number (1 or 2). This code is used in the further discussion. The factor genetic identity implies anatomical and physiological identity, and hence it is very reasonable to assume that the physiological and biomechanical properties of the vocal apparatus are rather similar in the monozygotic twin pairs and different in dizygotic pairs (cf. Lundstr€om, 1948; Langer et al., 1999; Kabban et al., 2001; Eguchi et al., 2004). However, various factors exist that can influence the anatomy and physiology involved in the speech production process, i.e., the use of a pacifier or a brace, dental operations, and habits like smoking or singing in a choir. The twins were asked about all of these possible influences, and no difference within any of the pairs appeared. Additional measurements of the height and weight of the participating twins support the assumption of a greater physiological resemblance in MZ than in DZ twins. However, because the size of the tongue and the shape of the palate are very relevant physiological factors regarding speech production (e.g., Brunner et al., 2009), further analysis of the palatal shapes (silicone dental palatal casts were created), and the tongue sizes were made and results are given in Sec. III A. 2. Environmental similarity of the subjects

As can be seen in Table I, the female twins do not differ with respect to the amount of time they have recently spent together. Both dizygotic twin pairs are still living together TABLE I. Overview of the twin pairs with information about the factors sex, age, genetic identity, amount of time spent together, and attitude toward being a twin (rated on a 5-point Likert scale from 1 “I don’t like being a twin” to 5 “I very much like being a twin”).

Twin MZf1 MZf2 MZm1 DZf1 DZf2

3770

Sex

Age

Genetic identity

Amount of time spent together

Attitude toward being a twin (1–5)

F F M F F

34 26 32 20 20

MZ MZ MZ DZ DZ

Nearly every day Live together Twice a month Live together Live together

5–5 5–5 5–5 4–5 5–5

J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

and have not been apart from each other longer than a few weeks. Of the monozygotic twin pairs, one female pair is also still living together, and the other female pair is living next door to each other and sees each other nearly every day. In contrast, the male pair is living separately from each other and sees each other only twice a month. The time the twins spend together can be considered an additional factor that might influence inter-speaker variability because the mutual influence of the twins and their shared social environment may play a role in shaping auditory goals. However, if childhood and adolescence are considered, which are the most crucial times in speech acquisition (Chambers, 2003; Clark, 2009), all twins are comparable. Thus a higher amount of inter-speaker variability in the DZ twin pairs than in the MZ twin pairs might reflect the higher amount of physiological diversity. Another important influencing factor to control for is the attitude toward being a twin and the attitude toward the sibling. Studies in social psychology dealing with speech accommodation and mimicry have shown that a positive correlation between the degree of accommodation and liking exists (cf. Chartrand and Bargh, 1999). If a person has a negative attitude toward being a twin, he or she might be more likely to separate himself/herself from his/her sibling and assert an individual style, which might also be mirrored in the speaking style and speech characteristics in general. Therefore separate interviews were conducted with each subject. When asked, all subjects tended to like being a twin and saw more advantages than disadvantages in being a twin. To quantify their statements, the twins were asked to make ratings on a 5-point Likert scale from 1 (“I do not like being a twin”) to 5 (“I very much like being a twin”). Number 3 served as a neutral position with no positive or negative attitude toward being a twin. All subjects showed a strong positive attitude toward being a twin (only ratings of 4 or 5; cf. Table I). Thus the factor attitude toward being a twin can be neglected because it should not influence the results as the pairs reveal no considerable differences in their feelings toward being a twin. Furthermore, the subjects were asked about their childhood as a twin, the amount of shared friends now and when they were younger, how they were brought up by the parents (if they dressed the twins always the same or put focus on an individual development of their children). All of these factors were the same within the participating subjects. 3. Speech material

The speech material investigated in the present study was obtained during a larger recording session with other target phonemes and carrier sentences (see Weirich, 2012 for a comprehensive overview of all analyzed data). For the present investigation, the looping pattern of the tongue back during /aCV/ sequences within the names “Haga,” “Hagu,” “Haka,” and “Haku” was chosen. The target words were part of the sentence “Ich gr€uße/wasche Haka/Haga/Haku/Hagu im Garten” (I greet/wash Haka/Haga/Haku/Hagu in the garden). During the recordings the subjects sat in a soundproof room, and the stimuli were presented on a screen that could Weirich et al.: Articulatory variability in twins’ speech

be seen through a window. The aim was to record ten repetitions of each speaker for each /aCV/ sequence, but because of technical problems during the articulatory recordings, not all repetitions could be used in the analysis. On average 9.4 (SD ¼ 1.1) repetitions for /aga/, 9 (SD ¼ 1.4) for agu, 9.6 (SD ¼ 0.97) for /aku/, and 9.8 (SD ¼ 0.63) for /aka/ could be used. B. Articulatory recordings and preprocessing

Electromagnetic articulography (Carstens AG100 EMA system) was used to record the motion of a coil placed on each speaker’s tongue back. Recordings were made at ZAS (Center for General Linguistics) in Berlin. Eight coils in total were attached to the subject’s tongue, lips, and jaw. Three coils were glued midsagittally to the tongue: One approximately 0.5 cm behind the tongue tip, one approximately 5 cm behind the tip on the tongue back, and a third one halfway in between the two, on the tongue dorsum. Another sensor was placed below the lower incisors to track jaw movements, and two further sensors were glued to the upper and the lower lip to record lip movements. Two sensors, one on the upper incisors and one on the bridge of the nose, served as reference sensors to compensate for head movements. The horizontal direction of each speaker’s bite plane was measured by letting the speakers bite on a plastic sheet on which two sensors were located in an anterior–posterior direction. After the recordings, the contour of the palate was recorded by moving a sensor along the palate from back to front. The articulatory data were preprocessed including correction algorithms for head movement, filtering of the data (low pass filter: Bandwidth of 18 Hz with a damping of 50 dB at 52 Hz), and rotation and translation of the position data (the supplementary correction program and preprocessing software that were used are described in more detail in Hoole, 1996). The sampling frequency of the processed articulatory data was 200 Hz. To investigate the degree of shared physiology between the twin pairs, silicone dental palatal casts (i.e., the negative of the dental impressions) were made for each speaker. An example of a pair of such casts is shown in Fig. 1 (right). Measurements were then taken to compare

the sizes of the palates: Maximal palatal height (measured as the distance between the base and highest point of the palatal cast), width (measured between the second molars), and palate length [measured midsagitally from the position of the second molars to the highest point on the palate before decreasing, see also Fig. 1 (right)]. Furthermore, the coil on the tongue back during the EMA recording had to be positioned at comparable points for the speakers within the twin pairs. Several precautions were taken to get data from coil positions that were as similar as possible. First, photographs were taken of the tongue with the glued coils on top of it and a ruler next to the tongue. Then a true-toscale template of the tongue with the coils of one of the twins was created with the help of the printed photograph. This template was then used as a reference for the second twin. The template, with holes at the coil positions of the first twin, was held on top of the tongue of the second twin, and the positions of the coils were marked through the holes of the template. Figure 1 (left) shows an example of the templates used. In addition, distances between the glued coils were measured and compared. Note though that potential differences in coil positions between speakers can be neglected with respect to the analysis of inter-speaker variability in the shape of the looping movements because curvature (that is relatively independent of coil positioning) was used for the comparisons. C. Segmentation of articulatory movements

The whole looping trajectory consists of a closing movement (from the vowel to the stop consonant) and an opening movement (from the stop consonant to the following vowel) of the tongue back. The start and end of the looping movement were marked by the vowel targets, which were determined by the lowest vertical position of the tongue back. The lowest vertical position is chosen because the aim was to compare the whole looping movement starting and ending with the vowel. Therefore we decided to use spatial landmarks that correspond to the lowest vertical position of the tongue back instead of landmarks corresponding to the tangential velocity. The acoustic recordings were used to verify the articulatory defined landmarks.

FIG. 1. Example of a true-to-scale tongue coil template with the three tongue coil positions (left) and the silicone palatal casts taken of DZf1 (right); lines indicate measurement points (horizontal line, palate width; vertical line, palate length). J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

Weirich et al.: Articulatory variability in twins’ speech

3771

To give a better impression of the analyzed data, Fig. 2 visualizes the articulatory movement of the tongue back during the sequence /aka/ for two renditions of one speaker (twin A from MZm1). The two different renditions of the speaker are indicated by different shades of gray, and the start of the movement is marked by an asterisk. It can be seen that the tongue does not simply lift for the /k/ and go down again for the /a/ but performs a forward loop. The tongue first lifts while moving backward and then slides forward along the palate during the occlusion of the /k/ before it moves down and back again to the second /a/.

TABLE II. Overview of the fixed factor (speaker) group with information on different levels and the number of pairs and comparisons. Number of comparisons

Factor (Speaker) group

Levels

Number of speaker-pairs

aga

agu

aka

aku

MZ

3

260

240

299

310

DZ Unrelated speaker (UN)

2 24

190 2162

170 1940

189 2398

310 2200

D. Data processing

To compare inter-speaker variability in articulation, the data had to be processed in several ways. First, to focus on the comparison of the shape of the looping movements between the speakers and to prevent an advantage of the potentially more similar coil positions in MZ twins than in DZ twins and unrelated speaker pairs, we calculated curvature (C) for each articulatory trajectory and took its absolute value (cf. Tasko and Westbury, 2004; O’Neill, 2006): C ¼ jðx00 y0  y00 x0 Þ=ðx02 þ y02 Þ3=2 j;

and (2) trajectories from two unrelated (sex matched) speakers. Third, the articulatory signals were aligned in time. This was achieved with the help of functional data analysis (see following text). The procedure of aligning was done for each speaker pair and each /aCV/ sequence separately. And last, a measurement of similarity in articulation that could be used for the statistical analysis (i.e., the mean sample-wise absolute difference between two aligned trajectories) was calculated for each pairwise comparison.

(1) 1. Functional data analysis

where y is the vertical and x is the horizontal position of the tongue back sensor, x0 and y0 are the velocities in the two directions, and x00 and y00 the accelerations.1 Following Tasko and Westbury (2004), the curvature data were natural log transformed and used for all further data processing and the comparison of the looping movements between the speakers. Second, trajectories were paired across speakers for each /aCV/ sequence in the following way: The set of ten tokens of one sequence by one speaker were compared to the set of ten tokens for the same sequence for another speaker. This comparison was done for all speaker pairs and all sequences. By doing so, we ended up with multiple pairwise comparisons (cf. Table II) that consist of (1) trajectories from two speakers of the same twin pair (MZ and DZ pairs)

FIG. 2. Articulatory movements of the tongue back (tback) in two /aka/ sequences of speaker A from MZm1 (different renditions marked in gray and black); a cross marks the beginning of the trajectory. 3772

J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

The technique of FDA to analyze time-varying signals was introduced by Ramsay and Silverman (1997), who use procrustean alignment to bring multiple signals into closer alignment with an average signal. Several studies have used this alignment method to study variability in speech: Lucero et al. (1997) used nonlinear temporal normalization to analyze the common pattern of lip movements and the variability in shape and timing. Lucero and Koenig (2000) investigated irregularities of voice signals, and Koenig et al. (2008) studied the intra-speaker variability in the fricative production of children vs adults by this means. Lucero and L€ofqvist (2005) examined articulatory variability in VCV sequences and found it to vary depending on the phonetic requirements of the consonant and the biomechanical characteristics of the articulatory structures involved. All of these studies illustrate that FDA is a powerful analytic tool for studying aspects of variability within the speech production process including articulation and acoustics. FDA alignment has several advantages over averaging conducted without time normalization (i.e., alignment) of the trajectories, where the average signal is computed by taking the average point by point. As a result, the averaged signal no longer resembles the individual waveforms due to cumulative distortions with increasing time. In linearly normalized averaging, the average is closer in shape to the original trajectories but still reveals significant differences in the amplitude. Here the major cause of distortion is phase variability: The interpolated signals can be slightly out of phase because of non-uniform timing changes (variation in the timing of landmarks) in the different trials. To remove phase variability, we adopt the registration method proposed by Lucero et al. (1997). First the time scales of the trajectories are linearly normalized so that all trajectories were set to a common length, and an average trajectory is obtained. Then warping Weirich et al.: Articulatory variability in twins’ speech

functions (non-linear transformations of the time scale) for each repetition are created in such a way that each aligned token (as a function of the transformed time) is close to the average. After a set of warping functions is obtained, a new average of the registered tokens is used to compute a new set of warping functions (cf. Lancia and Tiede, 2012 for a recent and more detailed overview of the method). In this study, FDA registration was conducted to get time-aligned articulatory trajectories that can be compared between the speakers.

unrelated speaker pairs. In addition, the potential impact of V2 (/a/ or /u/) and sonority (voiced or unvoiced stop) is investigated. A. Shared physiology in twins

III. RESULTS

Special emphasis was laid on the inspection and verification of a crucial assumption upon which this study is based: MZ and DZ twins are assumed to differ in their physiological resemblance, i.e., MZ twins are more similar in their physiology than DZ twins. Several of our results are in line with this hypothesis. First, the palatal contour was inspected. To facilitate the comparison between the speakers of one twin pair, the recorded palatal contours were aligned with the help of a MATLAB (version R2007a) script written by Pascal Perrier. The procedure finds vertical maxima of the palatal contours and aligns them. Then it looks for the vertical minima of the translated palate, computes the angle between the two alveolar regions, and rotates the transformed palate around the highest point to fit the position of the reference palate. This translation and rotation process was done for each twin pair, and Fig. 3 visualizes the result for MZf2. The palatal contours of the speakers A and B are plotted in gray and black, respectively. While the subplot on the left side shows the raw recorded data for both speakers, the subplot on the right side shows the original palatal contour for speaker B (black) but the adjusted palatal contour for speaker A (gray). In this way, as the figure shows, the shape of the palates could be compared in a better way despite potential intrinsic differences in skull morphology. Figure 4 displays the adjusted palatal contours for all twin pairs. The figure reveals nearly identical palatal shapes in all MZ twins but obvious differences in both DZ twins, especially in the steepness and length of the palatal rise. Similarly, the measurements of the silicone palatal casts also suggest that MZ twins are more similar than DZ twins: The difference in palate height was 0.1 cm for the male MZ twin pair, the female MZ twins did not differ at all in palate height, while both DZ twins did (DZf1: 0.1, DZf2: 0.3 cm). The difference in width was 0.1 cm for all three MZ pairs, 0.2 cm for DZf1, and 0.4 cm for DZf2. The difference in measured length of the palate reached 0.2 cm for MZf2 and only 0.1 cm for MZf1 and MZm1, but 0.6 cm for DZf1 and 0.5 cm for DZf2. In addition, it turned out that with DZf2, there were obvious differences in the length of the tongue. This became especially apparent when measuring the

In the following, first the amount of physiological resemblance between MZ and DZ twin pairs is discussed. Then a visual inspection of the looping movements and a quantitative analysis of the shape of the loops are presented. The raw data of the looping movements of the tongue back during /aka/ and /aku/ are shown for each speaker and compared within the twin pairs visually. After that mean distances that have been calculated between the aligned curvature data for each speaker pair and each /aCV/ sequence serve as an input for the quantitative analysis. A statistical model is estimated to analyze the differences in mean absolute distances between MZ speaker pairs, DZ speaker pairs, and

FIG. 3. Palatal contours of the two speakers of MZf2; different speakers are indicated by gray and black. The plot on the left side shows the raw data, the plot on the right side shows the rotated and translated data.

2. Multiple pairwise comparisons and absolute distances

The focus of this investigation lies on the comparison of articulatory movements between related speakers (twin pairs) and unrelated speakers. We aimed at ten repetitions for each of the ten speakers and each of the four sequences (/aga, aka, agi, aki/), but as mentioned in the preceding text because of technical problems, some data had to be excluded, and on average 9.4 (SD 1.0) trajectories for each speaker and sequence could be used for the comparisons (cf. Sec. II A 3). All possible pairwise combinations of the timealigned trajectories were constructed for each speaker pair and each sequence. The different speaker pairs can be separated into three groups according to their relationship to each other: The first group consists of speaker combinations within the MZ twin pairs (i.e., three twin pair comparisons). The second group consists of speaker combinations within DZ twin pairs (i.e., two twin pair comparisons). The third group consists of speaker combinations among unrelated speakers (i.e., 24 gender-matched unrelated speaker comparisons; cf. Table II). After that the mean absolute distance between the aligned curvature values for each comparison of two articulatory trajectories was calculated: The distance measure was calculated for each data point between the aligned curvature values and summed up for every possible comparison. In this way, a mean absolute distance in curvature for each comparison was obtained. The measured mean distances in curvature within the twin pairs and within the non-related pairs served as an index for the amount of interspeaker variability dependent on the different pairs. The resulting distances between the curvature values for the three groups MZ twins, DZ twins, and unrelated speaker pairs were used for the statistical analysis.

J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

Weirich et al.: Articulatory variability in twins’ speech

3773

FIG. 4. Aligned palatal contours of all twin pairs; different speakers of each twin pair are indicated by gray and black. The vertical line marks the highest point of the palate; the horizontal lines indicate the similarities or differences in the length of the palatal rise between the speakers of each pair.

distances between the three tongue coils. While all other pairs differed in the distance between the second (on the tongue mid) and the third coil (on the tongue back) by maximally 0.3 cm, the speakers of DZf2 varied by 0.6 cm pointing to a difference in tongue size. B. Visual inspection of the data

Within this section the shape of the loops is more closely inspected. The length and direction of the upward movement, the amount of horizontal sliding movement at the palate, and the amplitude and shape of the downward movement can be investigated. Here we will concentrate on the stimuli /aka/ and /aku/. Note that while for the following quantitative analysis all renditions of each speaker were taken into account, here an average loop for each speaker is shown. This was done for a better visualization of the general pattern of the loops. Figure 5 gives a closer look at the looping trajectories of the tongue back during /aka/ (above) and /aku/ (below) for the MZ twins. Each twin pair is shown in a subplot, and different speakers are indicated by gray and black. The start of the looping movement is marked by a cross. At first glance, much variability in the size and shape 3774

J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

of the loops can be seen between the pairs. However, within the pairs similarities are obvious. MZm1 differs from the other pairs with respect to the upward movement from the first vowel to the stop: Both speakers of MZm1 move upward and backward, while the other speakers start with a straight upward or even forward directed movement. This holds for /aka/ and /aku/. The speakers of MZm1 reveal in general quite similar shapes of the trajectories: The tongue moves upward and backward until it reaches the palate, then it slides forward and up along the palate until it moves down (for /aka/) and back (for /aku/), forming a triangle. The overall patterns of the loops are very similar between the brothers for both stimuli and point to an influence of shared physiology and biomechanics, especially in terms of consonant target and palatal contact. Both speakers of MZf2 reveal a rounder and smoother trajectory than the triangularly shaped loops of MZm1 and also—to a lesser degree—of MZf1. No clear start and end point of the palatal contact can be seen for any of the stimuli. While for /aku/ the downward movement to the second vowel is directed backward for both speakers, differences between the speakers can be found for /aka/: While twin A (gray) just lowers the tongue, twin B (black) also retracts it. The loops of the speakers of MFf1 Weirich et al.: Articulatory variability in twins’ speech

FIG. 5. Average looping trajectories of the tongue back during /aka/ (above) and /aku/ (below) for the monozygotic twin pairs (MZm1, MZf1, MZf2). Each twin pair is shown in a separate subplot. Different speakers are indicated by gray and black. Starting points of the trajectories are marked with a cross. y axis: Vertical movement of tongue back (in cm); x axis: Horizontal movement of tongue back (in cm).

are very similar for /aka/ but differ slightly for /aku/ in terms of the direction of the downward movement, which is straight for twin A (gray) and retracted for twin B (black). Figure 6 shows the looping trajectories of the dizygotic twin pairs (above /aka/, below /aku/). The form of the trajectory of DZf1 differs clearly for /aka/: While speaker A (gray) reveals a forward directed upward lifting of the tongue, a slightly horizontal movement at the palate and a downward and backward movement of the tongue to the second /a/, speaker B (black) shows first a forward and then backward directed lifting of the tongue with a very steep angle at the turning point at the palate and a downward and backward movement to the second vowel. No horizontal sliding at the palate can be seen. This twin pair also differs in the relative positions of the two vowels: While speaker A produces V2 at a more fronted position than V1, speaker B produces V2 at a more retracted position than V1. For /aku/, the difference is not that apparent. The speakers of DZf2 vary to some extent in the amount of horizontal movement at the palate. Speaker A (gray) reveals slightly more horizontal sliding movement than her sister. However, the general shape of the trajectory is quite similar. For /aku/, a difference can be seen: While speaker B (black) reveals a forward loop like all other speakers, speaker A (gray) does not show a loop at all but only a retraction to the second vowel.

aligned for each /aCV/ sequence separately. As an example, the time-aligned curvature data for the /aka/ sequence are shown for two MZ pairs and one DZ pair (Fig. 7). Each subplot shows curvature derived from approximately ten tokens per twin, hence 20 curvature lines per pair and subplot. As was mentioned in the preceding text, the magnitude of curvature increases for more sharply curved data. Differences can be seen between the pairs with respect to magnitude and number of peaks. While both speakers of MZf1 reveal one peak in the middle of the sequence, the speakers of MZm1 show two peaks. These differences in curvature mirror the differences described in the preceding text in the shape of the looping movements (cf. Figs. 5 and 6), in particular, in the nature of the turning point of the loop at the palate during the closure phase of the stop consonant. For example, the two peaks of MZm1 correspond to the two turning points during a looping trajectory: The first peak reflects the point where the tongue touches the palate, and the second peak reflects the point where the tongue lowers again for the second vowel. The lower curvature in between these two peaks reveals the straighter horizontal gliding at the palate. For DZf1, more variability can be seen between the curvature lines especially in the middle part of the trajectories. Some of the trajectories reveal a clear and strong peak, others are straighter. D. Quantitative comparison of the elliptical trajectories

C. Alignment of data

As described in the preceding text, all curvature data of speaker 1 and speaker 2 of a respective speaker pair were J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

Note that for the following statistical analysis, the logarithmic values (and not the linear values) of the calculated distances were used. This was done to normalize the Weirich et al.: Articulatory variability in twins’ speech

3775

FIG. 7. Time-aligned curvature signals for two MZ twin pairs and one DZ twin pair for /aka/; each graph shows approximately 20 curvatures lines, 10 for each twin of the pair indicated; y axis: Curvature; x axis: Normalized time. FIG. 6. Average looping trajectories of the tongue back during /aka/ (above) and /aku/ (below) for the dizygotic twin pairs (DZf1, DZf2). Each twin pair is shown in a separate subplot. Different speakers are indicated by gray and black. Starting points of the trajectories are marked by a cross. y axis: Vertical movement of tongue back (in cm); x axis: Horizontal movement of tongue back (in cm).

residuals, which is a mandatory assumption in linear mixed models (cf. Atkinson, 1985; Pinheiro and Bates, 2000). Figure 8 shows the boxplots of the log transformed distances for the three different groups and the four sequences. The figure reveals for all sequences the expected differences between the groups in the amount of mean absolute distances with MZ twin pairs showing the lowest values pointing to an influence of zygosity and hence physiology on speakerspecific differences in looping movements. Interestingly, no difference in mean values is apparent between DZ twins and unrelated speakers. To look for a significant effect of speaker pair (and in particular the pairs’ physiological similarity) on the similarity of the looping pattern, statistical tests were conducted using a linear mixed model (Pinheiro and Bates, 2000) as implemented in the lme4 package of the R software (version 2.14.1, R Development Core Team, 2008). A generalized linear mixed model was calculated with the measured pairwise mean absolute distance between the aligned curvature data as dependent variable. We included the fixed factors group (with the levels MZ, DZ, UN), vowel (levels /a/ and /u/), and voice (levels voiced and voiceless), and a pair specific random intercept for vowel and voice. The three different levels of the factor group mirror the different physiological similarities. Three pairs represent the group of the monozygotic twins (MZ), two pairs the dizygotic twins (DZ), and 24 different sex matched unrelated pairs were formed of the eight female speakers (UN, see Table II). The number of 3776

J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

comparisons for each speaker pair varies slightly between the pairs because the number of renditions (in total n ¼ 386) differs between the speakers (on average 9.4 per speaker and stimulus). Table II gives an overview of the fixed factor group with information on the levels, and the number of comparisons included in our model for each group and /aCV/ sequence (i.e., stimulus). Note that we used LMMs because

FIG. 8. Boxplots of logarithmic distances [log(distance)] separated by the three groups monozygotic twins (MZ), dizygotic twins (DZ), and unrelated speakers (UN). The median of the distribution in each group is visualized by filled dots; the boxes comprise 50% of the data; the whiskers extend to the most extreme data point which is no more than 1.5 times the interquartile range from the box; outliers are marked with open dots. Weirich et al.: Articulatory variability in twins’ speech

they are known to be robust to unbalanced data (cf. Baayen, 2008, p. 289) and our random effects structure was very conservative (allowing for speaker and stimuli specific variance terms). Moreover, the most important comparison in this study is the one between MZ and DZ twins, and here the groups have roughly the same size. Because it was hypothesized that physiology has an impact on the shape of the looping patterns, the levels of the factor group were ordered according to the speakers’ genetic and thus physiological similarity. It was hypothesized that the distances grow in the following way: (a) MZ pairs < (b) DZ pairs < (c) unrelated speakers (UN). Hence, the factor group was considered an ordered factor (MZ < DZ < UN) and was expressed through a successive difference contrast so that the significance of the difference between successive levels could be tested (Venables and Ripley, 2003). We also included the three-way-interaction term between group, vowel, and voice because a likelihood ratio test revealed a better fit of the model with the interaction term than without it (p < 0.001). After calculating the model, p values based on Markov-chain Monte Carlo sampling were computed. Table III gives a summary of the results of the fixed effects. Note that the given effect sizes are in logarithmic units, and the intercept is negative due to the log transformation of the distance measure. While the MZ twins reveal a significantly smaller distance than the DZ twins for the default level (which is /aga/, pMCMC < 0.001), the distances of the unrelated speaker pairs do not significantly differ from those of the DZ twin pairs. In general, productions with the vowel /a/ are more similar than productions with the vowel /u/ (pMCMC < 0.001). Furthermore, the model reveals a significant three-way-interaction between the group comparison MZ–DZ, vowel, and voice, which goes in the same direction as the MZ–DZ comparison of the default level, suggesting that the difference between MZ and DZ for the vowel /u/ with a voiceless consonant is just as big as the difference between MZ and DZ for /u/ with a voiced consonant. Note though that the interaction between the group comparison MZ–DZ and voice only marginally fails to show significance (p ¼ 0.053). The effect goes in the opposite direction with respect to the MZ–DZ comparison of the default level but

reaches only half of the strength. In other words, there is a tendency for a stronger effect of the group comparison between MZ and DZ twins in the voiced condition than in the voiceless condition for the vowel /a/. This tendency can also be seen in the size of the differences between the boxplots of Fig. 8 and might suggest a stronger influence of biomechanics in the voiced looping movements. We will come back to this in the discussion section. IV. DISCUSSION AND OUTLOOK

To sum up, articulatory variation was found between the different speakers in the analyzed looping trajectories. This was apparent already in the aligned curvature values for /aka/ (cf. Fig. 7) where differences appeared in the number and magnitude of curvature peaks. More inter-speaker variation was found within unrelated speaker pairs and DZ twin pairs than within MZ twin pairs. Results are thus in line with the assumption of an influence of biomechanics and physiology on the looping pattern in VCV sequences as suggested by, e.g., Perrier et al. (2003) and Birkholz et al. (2011). The visual inspection of the /aka/ and /aku/ sequences revealed more variation in DZ twins than in MZ twins (cf. Figs. 5 and 6): The MZ twins showed very similar looping patterns in terms of the size and the general shape of the loop as well as the direction of the upward movement (straight or back and up). The DZ twins revealed differences in the shape of the loop, the direction of the upward movement, and the amount of horizontal sliding movement at the palate. These findings were also supported by the quantitative analysis investigating the amount of mean absolute distance between the curvature signals: MZ and DZ twins differ significantly in their amount of articulatory inter-speaker variability in terms of the shape of the looping trajectories, while no difference is apparent between DZ twins and unrelated speakers. The factors vowel and voice revealed their significance only in terms of an influence of voice when the vowel /a/ was concerned: Here a tendency appeared for a bigger difference between MZ and DZ pairs in the voiced stimulus. However, for /u/, this effect was not apparent. This tendency can also be seen in the boxplots of Fig. 8. The

TABLE III. Summary statistics of the linear mixed model with mean absolute distance between the curvature comparisons as dependent variable, group [levels ¼ MZ, DZ, unrelated speaker (UN)], vowel (/a/, /u/), and voice (voiced, unvoiced) as fixed factors and pair (respective pair of comparison) as random factor (number of observation: 10 530). Interactions between factors marked by *.

(Intercept) Group (MZ vs DZ) Group (DZ vs UN) Vowel (a vs u) Voice (voiced vs voiceless) Group*vowel (MZ vs DZ * a vs u) Group*vowel (DZ vs UN * a vs u) Group*voice (MZ vs DZ * voiced vs voiceless) Group*voice (DZ vs UN * voiced vs voiceless) Vowel*voice (a vs u * voiced vs voiceless) Group*vowel*voice (MZ vs DZ * a vs u * voiced vs voiceless) Group*vowel*voice (DZ vs UN * a vs u * voiced vs voiceless)

J. Acoust. Soc. Am., Vol. 134, No. 5, November 2013

Effect size

t-value

pMCMC

1.335 0.808 0.246 0.423 0.014 0.368 0.054 0.439 0.178 0.075 0.458 0.056

13.48 2.78 1.05 3.76 0.17 1.11 0.20 1.89 0.95 1.80 3.75 0.57

Inter-speaker articulatory variability during vowel-consonant-vowel sequences in twins and unrelated speakers.

The purpose of this study is to examine and compare the amount of inter-speaker variability in the articulation of monozygotic twin pairs (MZ), dizygo...
901KB Sizes 0 Downloads 0 Views