Quantifying Voice Characteristics for Detecting Autism

Abstract

The presence of prosodic anomalies in autistic is recognized by experienced clinicians but their quantitative analysis is a cumbersome task beyond the scope of typical pen and pencil assessment. This paper proposes an automatic approach allowing to tease apart various aspects of prosodic abnormalities and to translate them into fine-grained, automated, and quantifiable measurements. Using a harmonic model (HM) of voiced signal, we isolated the harmonic content of speech and computed a set of quantities related to harmonic content. Employing these measures, along with standard speech measures such as loudness, we successfully trained machine learning models for distinguishing individuals with autism from those with typical development (TD). We evaluated our models empirically on a task of detecting autism on a sample of 118 youth (90 diagnosed with autism and 28 controls; mean age: 10.9 years) and demonstrated that these models perform significantly better than a chance model. Voice and speech analyses could be incorporated as novel outcome measures for treatment research and used for early detection of autism in preverbal infants or toddlers at risk of autism.

Combining voice and language features improves automated autism detection

“…In the present study, we aimed to confirm the performance of the automated voice and language measurements as previously established by Asgari et al (2021) and Salem et al (2021), and further, to evaluate any potential gains derived from their combination. Specifically, our approach is to estimate the accuracy of a combined voice and language model in correctly classifying ASD status in a mixed sample of already-diagnosed children and a non-autistic comparison group, approaching this work with a streamlined, purposeful inclusion of established voice and language measures.…”

“…More details about the study sample have been described previously by Salem et al (2021), where it was initially used. A subset of this sample was also used in Asgari et al (2021).…”

“…We analyzed 10 voice measures as described in Asgari et al (2021): Cepstrum, delta cepstrum, delta delta cepstrum, log spectral entropy, F0, jitter, shimmer, harmonic-to-noise ratio (HNR), H1H2, root mean square (RMS). Cepstrum is the shape of the spectral envelope that is extracted from cepstral coefficients.…”

“…Salem et al (2021) examined seven automated language measures (ALMs) and found them able to accurately classify already-diagnosed children into ASD and non-ASD groups even after controlling for age and IQ. Separately, we have also analyzed acoustic features of speech to evaluate their discriminant ability across diagnoses: Asgari, Chen, and Fombonne et al (2021) found that a harmonic model of speech could accurately differentiate between children with ASD and those who were typically developing (TD); again, results held true after further adjustment on IQ.…”

“…Our study had three specific aims. First, we wanted to replicate the findings of the voice model described by Asgari et al (2021) and of the language model described by Salem et al (2021), and compare their relative discriminant validity using the same methodology within the same sample. Second, we wanted to measure where gains in accuracy could be obtained by combining all features from both the voice and language models when compared to that of single modality models.…”