Acoustic-Prosodic and Physiological Response to Stressful Interactions in Children with Autism Spectrum Disorder

Full text

Acoustic-Prosodic and Physiological Response to Stressful Interactions
in Children with Autism Spectrum Disorder
Daniel Bone1, Julia Mertens2, Emily Zane2, Sungbok Lee1, Shrikanth Narayanan1, Ruth Grossman2,3
1Signal Analysis and Interpretation Laboratory (SAIL), USC, Los Angeles, CA, USA
2Face Lab, Emerson College, Boston, MA, USA
3Department of Communication Sciences and Disorders, Emerson College, Boston, MA, USA
dbone@usc.edu, http://sail.usc.edu
Abstract
Social anxiety is a prevalent condition affecting individuals to
varying degrees. Research on autism spectrum disorder (ASD),
a group of neurodevelopmental disorders marked by impair-
ments in social communication, has found that social anxiety
occurs more frequently in this population. Our study aims to
further understand the multimodal manifestation of social stress
for adolescents with ASD versus neurotypically developing
(TD) peers. We investigate this through objective measures of
speech behavior and physiology (mean heart rate) acquired dur-
ing three tasks: a low-stress conversation, a medium-stress in-
terview, and a high-stress presentation. Measurable differences
are found to exist for speech behavior and heart rate in relation
to task-induced stress. Additionally, we find the acoustic mea-
sures are particularly effective for distinguishing between diag-
nostic groups. Individuals with ASD produced higher prosodic
variability, agreeing with previous reports. Moreover, the most
informative features captured an individual’s vocal changes be-
tween low and high social-stress, suggesting an interaction be-
tween vocal production and social stressors in ASD.
Index Terms: stress, acoustic-prosody, physiology, autism
spectrum disorder, interaction
1. Introduction
Stressors are pervasive in our daily lives, impacting our mood,
our general sense of well-being, and even our health [1]. In fact,
our ability to deal with and adapt to stress is associated with
positive health outcomes. Anxiety disorders are the most preva-
lent disorder in the United States, affecting 18% of the popula-
tion [2]. When stress causes anxiety, it leads to increased physi-
ological arousal in the body [3], which we express in our verbal
and non-verbal behavior. One such stressor is social anxiety, as
with public speaking. Under stress, a person experiences un-
conscious sympathetic responses; e.g., the laryngeal folds may
tighten, leading to a rise in vocal pitch [4]. But there is still
much to learn about the ways in which individuals experience
and express stress; one viable approach uses scalable objective
measures of behavior, i.e., Behavioral Signal Processing [5].
A meta-analytic study reported that social anxiety occurs
more frequently in individuals with autism spectrum disorder,
or ASD [6], affecting 40% of the population. ASD is a highly
heterogeneous, highly prevalent (1 in 68 [7]) neurodevelopmen-
tal disorder defined by impairments in social communication
and reciprocity, as well as restricted, repetitive behavioral pat-
terns and interests [8]. Given the prevalence of anxiety in ASD,
researchers are striving to better understand when individuals
become stressed and how they respond (e.g., from skin conduc-
tance responses [9]). Since it can be difficult for those with
ASD to understand and communicate their emotions, acoustic
analyses may provide an effective measurement of stress.
There has been limited work specifically focused on the
acoustic correlates of stress, likely due to the challenges of
collecting high-quality, naturalistic speech under stress [10].
Speech researchers have primarily focused on optimizing emo-
tion classification within a database [11], whether the target is
categorical or dimensional (i.e., arousal, valence, dominance).
Yet, studies continue to find such models tuned for one database
do not readily transfer to another [12], which is critical to
the realization of speech-based behavioral health systems op-
erating “in the wild”. Approaches have included knowledge-
inspired system design [13], unsupervised neural-network adap-
tation [14], and multimodal behavioral integration [15, 16]. A
survey article by Juslin and Scherer reported several measures
that reliably increase with arousal or stress: pitch and intensity
mean and variability; the ratio of high-frequency energy; and
speaking rate [17]. In this study, we extract corresponding fea-
tures, but use functionals that were found to be more robust for
tracking arousal such as median or interquartile ratio [13].
The present work builds upon several of our previous stud-
ies which sought acoustic correlates of the “atypical prosody”
so commonly observed in autism spectrum disorders [18, 19,
20, 21, 22]. Our experiments [18, 19, 20] in a sample of 29
children from the USC CARE Corpus [23] found children with
increasing ASD severity spoke less, spoke slower, responded
later, had more variable prosody, and had more atypical voice
quality. Since atypicality is not universal in ASD, we have also
investigated human perception of atypicality or “awkwardness”.
We found that human agreement can be rather low for very spe-
cific dimensions of prosody, but that speech rate and rhythm
cues were highly predictive of overall perceived “awkardness”
in the read speech task [21]. In a large-scale study, we found
that prosodic variability was significantly higher for individu-
als with ASD compared to peers with non-ASD developmen-
tal disorders [22], aligning with previous findings in smaller
databases [24, 25]. Additionally, we presented novel features
that measured reduced coordination between pitch and intensity
or duration, quantifying a previous qualitative perception [26].
Previous work has primarily focused on a single modality;
in this novel study, we investigate the multimodal presentation
of stress in individuals with ASD and their neurotypically devel-
oping peers as they participate in a series of progressively stress-
ful interactions. As a measure of latent physiology, we consider
mean heart rate, which generally correlates with acute increases
in stress [27]. We also explore a set of acoustic-prosodic fea-
tures that are expected to be modulated by changes in affect [13]
as well as ASD symptoms [22], observing global tendencies as
well as changes that occur within a person between different
tasks. Through this study, we aim to enhance our understand-
ing of signal-derived measures of stress, which are crucial to
development of clinical engineering systems.

2. Methodology
In this section, we discuss: three interactions of varying stress
in our study; data collection and participant demographics;
acoustic-prosodic and physiological features related to stress
and autism; and data analysis and machine learning models.
2.1. Social Interactions of Variable Stress
Subjects participated in three types of social interactions ex-
pected to be progressively more stressful: a low stress one-on-
one conversation, a medium stress one-on-one interview, and
a high stress presentation to an audience. In the first task, the
subjects watched YouTube clips, and then discussed the clips
with a researcher (this conversation typically lasted under one
minute). Because the interaction was casual (participants were
not aware they were being recorded) and because the topic was
impersonal, we assume that individuals experienced low levels
of stress during these chats.
In the second scenario, the research assistant interviewed
the subject about their hobbies, family, and school for twenty
minutes (we analyze two minutes). Since questions were per-
sonal, the context was more formal, and the subject was aware
they were being recorded, we expected the subjects to feel an in-
creased level of pressure compared to the casual conversations.
The third interaction, an oral presentation, is hypothesized
to be the most stressful. Subjects were to develop the ending
of a story within five minutes, and then were asked to present
in front of a seated audience of three adult judges, video-edited
to appear as a live Skype call. Further intensifying any social
anxiety, the subjects were told that their performance would be
judged against their peers’ performances (see [9, 28]).
2.2. Data Collection and Participants
Experimental data consists of video-recorded interactions from
the three stressful scenarios for all subjects. Data are from 17
children with autism spectrum disorder (ASD) and 24 subjects
with neurotypical development (TD). Participant demographics
are presented in Table 1, including: ADOS diagnosis, age, and
number of audio samples for each of the three tasks.
Data were collected at a single site as part of an IRB-
approved study. Efforts were made to ensure video and audio
quality consistency between subjects and tasks. All recordings
took place in the same room. Camera microphone distance was
not constant across sessions, but the distance is not known to
be systematically different between groups. Still, we did not
feel confident in using voice quality measures, which were pre-
viously shown to be characteristic of ASD speech [19], with
the present far-field recordings; instead, we focus on prosodic
measures that may be more robust to any recording variability.
Table 1: Demographic information of all subjects presented as
mean (stdv.). Differences between ASD and TD subject’s age
and gender are non-significant (p>0.05).
N Age in yr. Female Acquired Task Audio
Low Medium High
ASD 17 13.7 (2.2) 19% 12 14 12
TD 24 13.4 (2.3) 39% 14 16 23
total 41 13.5 (2.2) 31% 26 30 34
Presence/absence of the subjects’ speech was manually an-
notated. Audio for several sessions was not available due to
recording difficulties or corrupted files. The number of session
for which audio features were extracted is displayed by task
in Table 1. Similar data loss occurred with heart-rate record-
ings. This loss primarily affects the joint audio-HR analyses,
for which missing HR data reduces data size by 19%.
2.3. Acoustic-Prosodic and Physiological Features
We computed five classes of features: segmental pitch cues;
segmental spectral cues; speaking rate; coordination between
prosodic modalities (a novel feature type from [22]); and heart
rate. Details of the feature extraction are provided below.
2.3.1. Speaking Rate
Because transcripts were not available for these data—with
which we could perform forced alignment—we needed to de-
termine syllabic boundaries directly from the audio signal.
Speaking rate estimation from prosodic and spectral signals
has been of some interest to the speech processing commu-
nity [29, 30, 31], but accurate estimation of syllabic boundaries
remains challenging. We implemented a version of a pitch-
and intensity-based method that has reported competitive per-
formance [31]. Visual inspection suggested this syllabic seg-
mentation was adequate. We computed two features using syl-
lable boundaries: median speaking rate (syl/s) and syllable du-
ration inter-quartile ratio (s), or IQR.
2.3.2. Segmental Prosodic Cues: Syllabic Contours
Pitch, volume, and the percentage of high-frequency energy are
all expected to increase with anxiety, stress, and arousal [17].
Further, segmental intonation that captures speaker idiosyn-
crasies in micro-prosodic production have been used to char-
acterize the speech of indivdiuals with ASD [19, 21, 22]. As
such, we compute nine segmental prosodic features from pitch
and intensity extracted via Praat [32], as well as median HF500
(the ratio of energy above 500Hz to that below) computed via
the vocal arousal score toolkit (VC-AS) [13].
In particular, we extracted syllable-level second-order poly-
nomial parametrization of pitch and intensity, then calculated
session-level medians and inter-quartile ratios of slope (four
features). The overall median and IQR of both log-pitch and
intensity are also calculated (four features). Aside from me-
dian log-pitch, all pitch analysis is performed in the OME (Oc-
tave MEdian) scale [33], a log-pitch transformation as in Eq. 1
through which speaker’s tend to have the same pitch range, i.e.,
one octave.
OME = log(f0Hz
)−log(median(f0Hz
)) (1)
Since a speaker’s range has been observed to reliably be one
OME around center in neutral speech, all speakers should have
a comparable range regardless of median pitch (unlike for Hz).
2.3.3. Prosodic Coordination Features
In previous work, we found that subjects with ASD showed re-
duced coordination of pitch with other prosodic markers [22].
Aside from ASD diagnosis, stress may affect this prosodic co-
ordination. Following the same approach [22], we quantified
the simultaneous movements of pitch, duration, and intensity
across syllables. These three feature streams are concatenated
per session, and then the Spearman’s rank-correlation coeffi-
cient is calculated pairwise, producing three features.
2.3.4. Physiological measure: heart rate
A person’s heart rate generally hastens under acute stress, and
has been specifically shown to increase in stressful speech inter-
actions [27]. We compute mean heart rate per session, while ex-
cluding sensor artifacts. Although heart-rate variability (HRV)
is commonly employed as a robust measure of complexity dif-
ferentially affected by acute versus chronic stress [27], compu-
tation requires a minimum sampling period of five minutes [34],
whereas each session lasts between 30 seconds and three min-
utes. Unlike HRV, mean HR is robust to the sampling period.

Table 2: Correlations of features with ADOS severity and best-estimate diagnosis. * indicates p<0.05; n.s. is non-significant.
Category Feature Task Stress Level ASD Diagnosis
Trend with task stress Sp. ρTrend with diagnosis Sp. ρ
Pitch cues
log-f0 median higher 0.49∗n.s. −0.15
log-f0 IQR n.s. −0.12 higher 0.33∗
log-f0 slope median n.s. 0.20 lower −0.27∗
log-f0 slope IQR n.s. 0.13 n.s. 0.14
Spectral cues
intensity median lower −0.47∗higher 0.29∗
intensity IQR lower −0.40∗higher 0.36∗
intensity slope median higher 0.56∗lower −0.22∗
intensity slope IQR n.s. −0.20 n.s. 0.18
HF500 median lower −0.28∗higher 0.32∗
Speaking Rate syllable rate median higher 0.23∗n.s. 0.10
syllable duration IQR lower −0.29∗higher 0.27∗
Prosodic
Coordination
corr. f0 & dur. less −0.29∗n.s. 0.05
corr. f0 & intensity less −0.25∗n.s. 0.14
corr. dur. & intensity less −0.21∗n.s. 0.01
Physiology heart rate median n.s. 0.20 n.s. 0.22
2.4. Statistical Analysis and Machine Learning
We conducted both statistical correlation analyses and clas-
sification experiments (with support vector machine via Lib-
linear software [35]). Parameters are tuned using two-level
nested cross-validation (CV), and averaged statistics of ten runs
of leave-one-subject-out CV are reported. Spearman’s rank-
correlation coefficient and unweighted average recall (UAR,
the mean of per-class recall) are selected as evaluation met-
rics. Note that in cases for which only two classes exists, the
p-value for Pearson’s rank-correlation coefficient is equivalent
to that from ANOVA; following, the same is true with Spear-
man’s rank-correlation coefficient, apart from the initial rank-
based feature transformation.
3. Results and Discussion
Relations between extracted behavioral features, task-induced
stress, and autism spectrum disorder (ASD) diagnosis can in-
form large-scale behavioral analyses. In Section 3.1, the objec-
tive speech and heart rate cues are analyzed versus the hypothe-
sized task-related stress level. Then, in Section 3.2, the cues are
used to differentiate task-type and to predict ASD diagnosis.
3.1. Correlational Feature Analysis
Acoustic-prosodic and heart rate feature correlations with task-
related stress level and ASD diagnosis are provided in Table 2.
The low stress (casual conversation), medium stress (interview),
and high stress (presentation) tasks are encoded with values of
0, 1, and 2, respectively, for purposes of analysis. Since the fea-
ture values are dependent on various sources of noise in data
collection and feature extraction processes, we cannot confi-
dently state that a construct is or is not informative of a target
variable, only whether an extracted feature is in this experiment.
Segmental pitch cues capture short-term tendencies in us-
age of fundamental frequency. We expected median pitch to
shift upward with increasing stress for a given speaker [17]; in
fact, the only significant relation between pitch cues and task
stress-level is that speakers tend to increase their pitch in more
stressful tasks (p<0.05). Although one may expect pitch vari-
ability to have also increased, median pitch may be more robust,
as it has been shown for a related percept, vocal arousal [13].
Regarding ASD, children with higher social-communicative
deficits have previously shown more negative pitch curvature—
which is possibly perceived as “flat” or “monotone” [18]—and
displayed more prosodic variability [22]. For both cases, we
confirm the previous findings; i.e., log-f0 variability is higher
for ASD subjects, while log-f0 slope is lower. The relative
(person-specific) changes in log-f0 between tasks are not signif-
icantly different between groups, while for both ASD and TD
subjects there is a significant increase in log-f0 between low and
high stress tasks (Figure 1a).
Segmental intensity cues may be similarly influenced by
stress; however, sound level is also a function of distance and
angle to the microphone. As such, we should interpret the in-
tensity findings cautiously. Contrary to expectations, we find
med_stress_f0 − low_stress_f0 high_stress_f0 − low_stress_f0
−0.2
0
0.2
0.4
0.6
0.8
f0 change (fraction)
ASD TD
(a) Relative fundamental frequency changes between interactions.
med_stress_HR − low_stress_HR high_stress_HR − low_stress_HR
−5
0
5
10
15
20
HR change (beats/min)
ASD TD
(b) Heart rate (HR) changes (beats/min) between interactions.
Figure 1: Mean relative changes (and standard-deviation) between stressful interactions: low (conversation), medium (interview), and
high (presentation). For both groups and features, changes between low and high stress scenarios are statistically significant (p<0.05).

that intensity median and IQR tend to decrease with increas-
ing task stress, as does HF500 (which is a primary correlate of
arousal [13]). Given the variability in audio conditions, it is dif-
ficult to ascertain if an incidental result of microphone distance
is being measured, or if, in fact, subjects reduce their volume
given increasing stress. We also observe that ASD subjects have
higher and more variable vocal intensity.
Speaking rate is another reported correlate of perceived vo-
cal arousal [17]. In our data, subjects in higher stress situations
speak faster, but also with lower durational variability; this in-
dicates a more rigid, tense speech production. Also, children
with ASD spoke with more durational variability—yet another
indicator of increased variability associated with ASD.
Following our previous quantitative support [22] of a qual-
itative finding [26, 36], we suspected that ASD subjects with
“atypical” prosody were sometimes modulating pitch incongru-
ously with other modalities. We quantified this prosodic coor-
dination as the pairwise correlation between three modalities:
syllabic fundamental frequency, vocal intensity, and duration.
In this study, we found no statistical difference between ASD
and TD groups. However, we did find that subjects in higher
stress tasks tended to have less coordination between prosodic
modalities–a possible result of reduced motor control as a phys-
iological response to stress.
Lastly, we investigate our single physiological measure,
mean heart rate, which is anticipated to increase with task stress.
We find that overall heart rate was not significantly higher in
the higher stress tasks, and that there is no relation with diag-
nosis. But because heart rate varies from person to person (due
to general health, respiration rate, etc.) we calculated relative
(per-person) increases between tasks; in fact, we find that mean
HR increased from low stress to high stress tasks (Figure 1b).
3.2. Prediction Experiments
Machine learning allows for building systems that incorporate
multivariate dependencies which are not obvious in statistical
observation. In this section, we analyze the performance of dif-
ferent feature categories for predicting tasks of varying stress
(Table 3) and for predicting ASD diagnosis (Table 4). In ad-
dition to session-level features, we introduce relative features
(as in Figure 1), which measure intra-personal changes between
tasks. We compute relative changes between five features (log-
f0, intensity, HF500, speaking rate, and HR) for all three task
comparisons (medium-low, high-medium, and high-low).
We initially examine the predictive power of acoustic and
heart rate features across diagnostic groups for task-stress as
shown in Table 3. We report both UAR and Spearman’s rank-
correlation coefficient, given that the task stress-labels are ordi-
nal. The acoustic features are significantly predictive of ASD
severity within both ASD and TD populations (p<0.05). This
is an intuitive finding, given the theoretical underpinnings and
empirical evidence for the relation between the acoustic features
and stress/anxiety/arousal. Interestingly, heart rate level alone
is only predictive of stress level for the ASD subjects. As stated
Table 3: Classification of task stress level from acoustic-
prosodic and HR features. Results are presented in terms of
UAR (baseline=33%) and Spearman’s rank-correlation coeffi-
cient. Bolded statistics are significant at the α=0.05 level.
Features
Group Acoustic Heart Rate Combined
ASD 56% 0.59 54% 0.45 62% 0.70
TD 52% 0.57 34% 0.10 55% 0.60
All 70% 0.72 41% 0.17 67% 0.69
Table 4: Classification of ASD diagnosis from acoustic-
prosodic and HR features in different stressful interactions. Re-
sults are presented in terms of UAR (baseline=50%). Bolded
statistics are significant at the α=0.05 level. ”Session” refers
to an individual task, while ”relative” refers to comparisons be-
tween low/medium, medium/high, and low/high, respectively.
Features
Acoustic Heart rate Combined
Task Session Relative Session All
Low 70 64 (M-L) 54 65
Medium 73 73 (H-M) 56 77
High 70 87 (H-L) 57 84
All 69 75 (All) 61 73
previously, dependence of resting heart rate on external factors
may overcome the influence of certain acute stressors; thus, the
relative HR features are most appropriate and useful. Feature
fusion generally leads to nominal increases in performance.
Next we consider classification of ASD diagnosis with be-
havioral features as a function of stressful interaction type. We
hypothesized that there may be differences in the ways in which
individuals on the spectrum experience and express stress in
comparison to their typically developing peers. However, it
is unclear if such a difference exists. We do observe that for
the combined feature set classification performance is higher
for more stressful tasks (65%, 75%, and 84%, respectively);
but this appears driven primarily by the relative-change acoustic
features (i.e., 64%, 73%, and 87%, respectively). Thus, further
investigation is required to ascertain the degree to which stress
changes modulate vocal behavior in ASD. Our physiological
measure, task-level mean heart rate, did not achieve significant
prediction, nor did the relative heart rate features of Table 3 (not
shown due to space constraints).
The most informative features are certainly the relative
acoustic features, wherein vocal changes between low and high
stress tasks achieve a predictive performance of 87% UAR.
Session-specific features achieve a lower performance, although
one that is consistent across tasks. This highlights an important
concept, that each person has their own baseline, and the way in
which behavior deviates from that baseline is quite informative.
4. Conclusion
In this work, we examined vocal and physiological measures
of stress during social interactions designed to induce varying
levels of anxiety in individuals with autism spectrum disorder
and typically developing peers. Certain findings corroborate
previous reports regarding acoustic-prosodic markers of autis-
tic speech, including increased prosodic variability (pitch, in-
tensity, and speaking rate) and more negative pitch slope (a pos-
sible correlate of perceived “monotone” speech in ASD). Fur-
thermore, measurable differences in behavioral features were
demonstrated through classification experiments in which those
features could identify the corresponding stressful task as well
as diagnosis. It is compelling that intra-personal acoustic devi-
ation between low and high stress tasks was quite informative
of ASD diagnosis. Still, further investigation is needed to better
understand the covariation of covert and overt behavioral cues,
acute stress, and autism spectrum disorder.
5. Acknowledgments
This work was supported by funds from the National Science
Foundation and the National Institute of Health. We thank the
children and families who generously gave their time.

6. References
[1] N. Schneiderman, G. Ironson, and S. D. Siegel, “Stress and health:
psychological, behavioral, and biological determinants,” Annu.
Rev. Clin. Psychol., vol. 1, pp. 607–628, 2005.
[2] R. C. Kessler, W. T. Chiu, O. Demler, and E. E. Walters, “Preva-
lence, severity, and comorbidity of 12-month dsm-iv disorders in
the national comorbidity survey replication,” Archives of general
psychiatry, vol. 62, no. 6, pp. 617–627, 2005.
[3] T. Steimer, “The biology of fear-and anxiety-related behaviors,”
Dialogues in clinical neuroscience, vol. 4, pp. 231–250, 2002.
[4] K. R. Scherer, “Vocal affect expression: a review and a model for
future research.” Psychological bulletin, vol. 99, no. 2, 1986.
[5] S. Narayanan and P. G. Georgiou, “Behavioral signal process-
ing: Deriving human behavioral informatics from speech and lan-
guage,” Proceedings of the IEEE, vol. PP, no. 99, pp. 1–31, 2013.
[6] F. J. van Steensel, S. M. B¨
ogels, and S. Perrin, “Anxiety disor-
ders in children and adolescents with autistic spectrum disorders:
a meta-analysis,” Clinical child and family psychology review,
vol. 14, no. 3, p. 302, 2011.
[7] J. Baio, “Prevalence of autism spectrum disorder among children
aged 8 years-autism and developmental disabilities monitoring
network, 11 sites, united states, 2010.” Morbidity and mortality
weekly report. Surveillance summaries, vol. 63, no. 2, p. 1, 2014.
[8] A. P. Association et al.,Diagnostic and statistical manual of men-
tal disorders, (DSM-5®). American Psychiatric Pub, 2013.
[9] J. Mertens, E. Zane, K. Neumeyer, and R. Grossman, “How anx-
ious do you think i am? relationship between state and trait anxi-
ety in children with and without asd during social tasks,” Journal
of Autism and Developmental Disorders, pp. 1–12, 2017.
[10] J. H. Hansen, S. E. Bou-Ghazale, R. Sarikaya, and B. Pellom,
“Getting started with susas: a speech under simulated and actual
stress database.” in Eurospeech, vol. 97, no. 4, 1997, pp. 1743–46.
[11] C.-C. Lee, E. Mower, C. Busso, S. Lee, and S. Narayanan,
“Emotion recognition using a hierarchical binary decision tree ap-
proach,” Speech Comm., vol. 53, no. 9, pp. 1162–1171, 2011.
[12] B. Schuller, B. Vlasenko, F. Eyben, M. Wollmer, A. Stuhlsatz,
A. Wendemuth, and G. Rigoll, “Cross-corpus acoustic emotion
recognition: Variances and strategies,” IEEE Transactions on Af-
fective Computing, vol. 1, no. 2, pp. 119–131, 2010.
[13] D. Bone, C.-C. Lee, and S. Narayanan, “Robust unsupervised
arousal rating: A rule-based framework with knowledge-inspired
vocal features,” IEEE Transactions on Affective Computing,
vol. 5, no. 1, pp. 201–213, 2014.
[14] J. Deng, Z. Zhang, F. Eyben, and B. Schuller, “Autoencoder-based
unsupervised domain adaptation for speech emotion recognition,”
IEEE Signal Proc. Letters, vol. 21, no. 9, pp. 1068–1072, 2014.
[15] C. Busso, Z. Deng, S. Yildirim, M. Bulut, C. M. Lee,
A. Kazemzadeh, S. Lee, U. Neumann, and S. Narayanan, “Anal-
ysis of emotion recognition using facial expressions, speech and
multimodal information,” in Proceedings of the 6th international
conference on Multimodal interfaces. ACM, 2004, pp. 205–211.
[16] M. Valstar, J. Gratch, B. Schuller, F. Ringeval, D. Lalanne, M. Tor-
res Torres, S. Scherer, G. Stratou, R. Cowie, and M. Pantic, “Avec
2016: Depression, mood, and emotion recognition workshop and
challenge,” in Proceedings of the 6th International Workshop on
Audio/Visual Emotion Challenge. ACM, 2016, pp. 3–10.
[17] P. Juslin and K. Scherer, The New Handbook of Methods in Non-
verbal Behavior Research. Oxford: Oxford University Press.,
2005, ch. 3. Vocal Expression of Affect, pp. 65–135.
[18] D. Bone, M. P. Black, C.-C. Lee, M. E. Williams, P. Levitt, S. Lee,
and S. Narayanan, “Spontaneous-Speech Acoustic-Prosodic Fea-
tures of Children with Autism and the Interacting Psychologist.”
in Proceedings of Interspeech, 2012, pp. 1043–1046.
[19] ——, “The Psychologist as an Interlocutor in Autism Spectrum
Disorder Assessment: Insights from a Study of Spontaneous
Prosody,” Journal of Speech, Language, and Hearing Research,
vol. 57, pp. 1162–1177, 2014.
[20] D. Bone, C.-C. Lee, T. Chaspari, M. Black, M. Williams, S. Lee,
P. Levitt, and S. Narayanan, “Acoustic-prosodic, turn-taking, and
language cues in child-psychologist interactions for varying social
demand,” in Proceedings of Interspeech, 2013.
[21] D. Bone, M. P. Black, A. Ramakrishna, R. Grossman,
and S. Narayanan, “Acoustic-prosodic correlates of ‘awkward’
prosody in story retellings from adolescents with autism,” in Pro-
ceedings of Interspeech, 2015.
[22] D. Bone, S. Bishop, R. Gupta, S. Lee, and S. Narayanan,
“Acoustic-prosodic and turn-taking features in interactions with
children with neurodevelopmental disorders,” Proceedings of In-
terspeech, pp. 1185–1189, 2016.
[23] M. P. Black, D. Bone, M. E. Williams, P. Gorrindo, P. Levitt, and
S. S. Narayanan, “The USC CARE Corpus: Child-Psychologist
Interactions of Children with Autism Spectrum Disorders,” in
Proceedings of Interspeech, 2011.
[24] J. J. Diehl, D. Watson, L. Bennetto, J. McDonough, and C. Gun-
logson, “An Acoustic Analysis of Prosody in High-Functioning
Autism,” Applied Psycholinguistics, vol. 30, pp. 385–404, 2009.
[25] R. B. Grossman, L. R. Edelson, and H. Tager-Flusberg, “Emo-
tional facial and vocal expressions during story retelling by chil-
dren and adolescents with high-functioning autism,” Journal of
Speech, Language, and Hearing Research, vol. 56, no. 3, pp.
1035–1044, 2013.
[26] C. Baltaxe, J. Q. Simmons, and E. Zee, “Intonation patterns in
normal, autistic and aphasic children,” in Proceedings of the Tenth
International Congress of Phonetic Sciences. Foris Publications
Dordrecht, The Netherlands, 1984, pp. 713–718.
[27] C. Schubert, M. Lambertz, R. Nelesen, W. Bardwell, J.-B. Choi,
and J. Dimsdale, “Effects of stress on heart rate complexity?a
comparison between short-term and chronic stress,” Biological
psychology, vol. 80, no. 3, pp. 325–332, 2009.
[28] C. Kirschbaum, K.-M. Pirke, and D. H. Hellhammer, “The ‘trier
social stress test’–a tool for investigating psychobiological stress
responses in a laboratory setting,” Neuropsychobiology, vol. 28,
no. 1-2, pp. 76–81, 1993.
[29] N. Morgan and E. Fosler-Lussier, “Combining multiple estimators
of speaking rate,” in ICASSP, vol. 2. IEEE, 1998, pp. 729–732.
[30] D. Wang and S. S. Narayanan, “Robust speech rate estimation for
spontaneous speech,” IEEE Transactions on Audio, Speech, and
Language Processing, vol. 15, no. 8, pp. 2190–2201, 2007.
[31] N. H. De Jong and T. Wempe, “Praat script to detect syllable nu-
clei and measure speech rate automatically,” Behavior research
methods, vol. 41, no. 2, pp. 385–390, 2009.
[32] P. Boersma, “Praat, a system for doing phonetics by computer,”
Glot International, vol. 5, no. 9/10, pp. 341–345, 2001.
[33] C. De Looze and D. Hirst, “The ome (octave-median) scale: A
natural scale for speech prosody,” in Proceedings of the 7th Inter-
national Conference on Speech Prosody (SP7), 2014.
[34] T. F. of the European Society of Cardiology et al., “Heart rate
variability standards of measurement, physiological interpreta-
tion, and clinical use,” Eur heart J, vol. 17, pp. 354–381, 1996.
[35] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin,
“Liblinear: A library for large linear classification,” The Journal
of Machine Learning Research, vol. 9, pp. 1871–1874, 2008.
[36] L. D. Shriberg, R. Paul, J. L. McSweeny, A. Klin, D. J. Cohen,
and F. R. Volkmar, “Speech and Prosody Characteristics of Ado-
lescents and Adults with High-Functioning Autism and Asperger
Syndrome,” JSLHR, vol. 44, pp. 1097–1115, 2001.