Voice Patterns in Children with Autism Spectrum Disorder: Predicting Diagnostic Status and Symptoms Severity

Background: Individuals with autism spectrum disorder (ASD) tend to have atypical modulation of speech, often described as awkward, monotone, or sing-songy (Shriberg et al., 2001). The patterns may be one of the most robust signals of social communication deficits in ASD (Paul et al., 2005).  However, it has proven difficult to determine a consistent set of acoustic features that can account for these perceived differences. Using Recurrence Quantification analysis of acoustic features, Fusaroli et al. (Fusaroli, Bang, & Weed, 2013) demonstrated a high efficacy of identifying voice patterns characteristic of adult Danish speakers with Asperger’s syndrome.

Objectives: We systematically quantify and explore speech patterns in Danish children (8-12 years) with and without autism. We employ traditional and non-linear techniques measuring the structure (regularity and complexity) of speech behavior (i.e. fundamental frequency, use of pauses, speech rate). Our aims are (1) to achieve a more fine-grained understanding of the speech patterns in children with ASD, and (2) to employ the results in a supervised machine-learning process to determine whether acoustic features can be used to predict diagnostic status and severity of the symptoms.

Methods: Our analysis was based on previously-acquired repeated narratives (TOMAL-2 (Reynolds & Voress, 2007)). We tested 25 Danish children diagnosed with ASD and matched controls. Participants had been diagnosed using ADOS and ADI-R and their symptoms assess with SRS and SCQ. Transcripts were time-coded, and pitch (F0), speech-pause sequences and speech rate were automatically extracted. Per each prosodic feature we calculated traditional statistical measures. We then extracted non-linear measure of recurrence: treating voice as a dynamical system, we reconstructed its phase space and measured the number, duration and structure of repeated trajectories in that space(Marwan, Carmen Romano, Thiel, & Kurths, 2007). The results were employed to train (1) a linear discriminant function algorithm to classify the descriptions as belonging either to the ASD or the control group, and (2) a multiple linear regression to predict scores in Social Responsiveness Scale (SRS) and Social Communication Questionnaire (SCQ). Both models were developed and tested using 1000 iterations of 10-fold cross-validation (to test the generalizability of the accuracy) and variational Bayesian mixed-effects inferences (to compensate for biases in sample sizes).

Results: While traditional measured did not allow for accurate classification, recurrence measures allowed to define voices as autistic or not with balanced accuracy > 77% (p<.00001, CI =71.79%- 81.01%), sensitivity: 79.19%, specificity: 82.37%. Recurrence also allowed to explain variance in the severity of the symptoms: 42.76% (p<.00001) for SCQ and 55.80% for SRS (p<.00001, 48.18% for Social Consciousness, 53.92% for Social Cognition, 54.46% for Social Communication, 47.18% for Social Motivation and 61,04% for Autistic Mannerism). Autistic voice can be characterized as more regular (i.e. with regularly repeated patterns) pitch and pause use than neurotypical voices,  

Conclusions: Non-linear time series analyses techniques suggest that there are quantifiable acoustic features in speech production of children with ASD that both distinguish them from typically developing speakers and reflect the severity of the symptoms.