Prosodic signatures of ASD severity and developmental delay in preschoolers

Abstract

Atypical prosody in speech production is a core feature of Autism Spectrum Disorder (ASD) that can impact everyday life communication. Because the ability to modulate prosody develops around the age of speech acquisition, it might be affected by ASD symptoms and developmental delays that emerge at the same period. Here, we investigated the existence of a prosodic signature of developmental level and ASD symptom severity in a sample of 74 autistic preschoolers. We first developed an original diarization pipeline to extract preschoolers' vocalizations from recordings of naturalistic social interactions. Using this novel approach, we then found a robust voice quality signature of ASD developmental difficulties in preschoolers. Furthermore, some prosodic measures were associated with one year later outcome in participants who had not acquired speech yet. Altogether, our results highlight the potential benefits of automatized diarization algorithms and prosodic metrics for digital phenotyping in psychiatry, helping clinicians establish early diagnosis and prognosis.

Fig. 1. Prosody in autistic preschoolers with and without phrase speech.

A Partial Least Square Correlation (PLSC) in preschoolers with ASD without phrase speech (PreSp, n = 42). PLSC explored the association between behavioral (ADOS CSS in blue and DQ in pink) and prosodic measures relating to intonation (blue), loudness (purple), quality of voice (pink) and rhythm (green). Behavioral and prosodic saliencies of the first principal component are displayed. Error bars represent 5–95th percentiles of 500 bootstrap samples, and robust results are highlighted in yellow. B PLSC in autistic preschoolers with phrase speech (Sp, n = 32). ADOS Autism Diagnostic Observation Schedule, CSS Calibrated Severity Score, DQ Developmental Quotient, RRB Repetitive and Restricted Behaviors.

Fig. 2. Association between prosody and language production stage.

Significant correlation component of Partial Least Square Correlation (PLSC) in 74 preschoolers with ASD is displayed. PLSC explored the association between behavioral contrast (Sp versus PreSp) and prosodic measures relating to intonation (blue), loudness (purple), quality of voice (pink) and rhythm (green). Behavioral contrast and prosodic saliencies of the first principal component are displayed. Error bars represent 5–95th percentiles of 500 bootstrap samples, and robust results are highlighted in yellow. PreSp: without phrase speech; Sp: with phrase speech.

Fig. 3. Associations between prosody and subsequent behavioral rates of change.

PLSC in a subsample (n = 32) of preschoolers without phrase speech (PreSp). PLSC explored the association between the prosodic measures at the first visit and the behavioral rates of change (SPC) in ADOS CSS (in blue) and DQ (in pink) during the following year. Prosodic measures comprised intonation (blue), loudness (purple), quality of voice (pink) and rhythm (green). Behavioral and prosodic saliencies of the first principal component are displayed. Error bars represent 5–95th percentiles of 500 bootstrap samples, and robust results are highlighted in yellow. ADOS Autism Diagnostic Observation Schedule, CSS Calibrated Severity Score, DQ Developmental Quotient, RRB Repetitive and Restricted Behaviors, SPC Symmetrized Percent Change.

Fig. 4. Summary of the preprocessing pipeline.

A Recording of the ADOS administered by a male examiner with either a Shure© MX202 Microflex® overhead microphone or the GoPro© Hero7 built-in microphone. B After pre-emphasis filtering and denoising (not illustrated), a voice detection algorithm was applied on the complete audio track. Noise was removed and voiced segments were concatenated. C Then, we applied our original diarization algorithm to automatically identify, extract and concatenate the segments of children vocalizations. The resulting concatenated audio track was carefully inspected and edited by an examiner (not illustrated). D Ultimately, a set of prosodic parameters were extracted from the isolated children vocal productions using the GeMAP pipeline. Illustration of prosodic features (above right) was obtained using WaveSurfer.js for illustration purposes. Here, we display pitch contour (below) and formants trajectories on a spectrogram (above) of a 3 s excerpt from a single participant. Informed consent for the photograph displayed was signed by the participant’s parents. F0 relates to fundamental frequency (or pitch) and F1, F2 and F3 relate to first, second and third formants respectively. ADOS Autism Diagnosis Observation Schedule.