Voice acoustics allow classifying autism spectrum disorder with high accuracy

Abstract

Early identification of children on the autism spectrum is crucial for early intervention with long-term positive effects on symptoms and skills. The need for improved objective autism detection tools is emphasized by the poor diagnostic power in current tools. Here, we aim to evaluate the classification performance of acoustic features of the voice in children with autism spectrum disorder (ASD) with respect to a heterogeneous control group (composed of neurotypical children, children with Developmental Language Disorder [DLD] and children with sensorineural hearing loss with Cochlear Implant [CI]). This retrospective diagnostic study was conducted at the Child Psychiatry Unit of Tours University Hospital (France). A total of 108 children, including 38 diagnosed with ASD (8.5 ± 0.25 years), 24 typically developing (TD; 8.2 ± 0.32 years) and 46 children with atypical development (DLD and CI; 7.9 ± 0.36 years) were enrolled in our studies. The acoustic properties of speech samples produced by children in the context of a nonword repetition task were measured. We used a Monte Carlo cross-validation with an ROC (Receiving Operator Characteristic) supervised k-Means clustering algorithm to develop a classification model that can differentially classify a child with an unknown disorder. We showed that voice acoustics classified autism diagnosis with an overall accuracy of 91% [CI95%, 90.40%-91.65%] against TD children, and of 85% [CI95%, 84.5%–86.6%] against an heterogenous group of non-autistic children. Accuracy reported here with multivariate analysis combined with Monte Carlo cross-validation is higher than in previous studies. Our findings demonstrate that easy-to-measure voice acoustic parameters could be used as a diagnostic aid tool, specific to ASD.

Reliably quantifying the severity of social symptoms in children with autism using ASDSpeech

“…According to these studies, ASD individuals exhibit atypical speech characteristics, including significantly fewer speech sounds [13], fewer phonemes per utterance [11], and fewer conversational turns [13] as well as significantly larger mean pitch [9], [19], larger pitch range [8], [9], and larger pitch variability [8], [9] than TD children. Differences in these and other speech characteristics have enabled the development of classification algorithms that can separate ASD and TD individuals with reported accuracy rates as high as 75-98% [17]- [23]. However, it is important to note that most classification studies have included small samples with <40 ASD participants per study who were recorded at different ages and in different contexts ranging from 12-minute parent-child free play sessions [17] to repetition of non-word syllable groups [23].…”

“…Differences in these and other speech characteristics have enabled the development of classification algorithms that can separate ASD and TD individuals with reported accuracy rates as high as 75-98% [17]- [23]. However, it is important to note that most classification studies have included small samples with <40 ASD participants per study who were recorded at different ages and in different contexts ranging from 12-minute parent-child free play sessions [17] to repetition of non-word syllable groups [23]. Identifying the specific age-dependent speech characteristics that reliably distinguish between ASD and TD children will require additional validation with considerably larger datasets that capture broad ASD heterogeneity.…”

“…According to these studies, ASD individuals exhibit atypical speech characteristics, including significantly fewer phonemes per utterance 11 , fewer conversational turns 13 , higher pitch 9,19 , and larger pitch range and variability 8,9 than TD children. Differences in these and other speech characteristics have enabled the development of machine and deep learning classification algorithms that can identify ASD and TD individuals with reported accuracy rates of 75-98% [17][18][19][20][21][22][23] .…”

“…Additional studies have trained machine learning and deep learning algorithms to classify ASD and TD children based on extracted speech features [17]- [24] and have reported classification accuracies of 75% to 98% [19]. Different studies used different speech characteristics, including linguistic features such as vocabulary and fluency [24] and acoustic features such as pitch [18]- [20], [22], [23], jitter [20], [23], shimmer [20], [23], energy [18], [19], Zero-Crossing Rate (ZCR) [18], [19], Mel-Frequency Cepstral Coefficients (MFCCs) [19], Linear Predictive Coefficients (LPCs) [19], formants [18], [19], [23], speech rate [22], and the extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS) (88 acoustic features commonly used in speech analyses) [17], [21], [22]. Some of these studies utilized classical machine learning techniques, such as Support Vector Machine (SVM) classification [18], [20], [22], [24] or K-Nearest Neighbors (KNN) and K-Means clustering [18], [23] while others used Probabilistic Neural Network (PNN) [19] and Bidirectional Long Short-Term Memory (BLSTM) [17] deep learning algorithms.…”

“…Several studies have used automated speech analysis techniques to classify ASD and TD children based on extracted speech features [17][18][19][20][21][22][23][24] . In some studies, diagnostic classification was based on linguistic features such as vocabulary and fluency 24 while in others it was based on acoustic features such as pitch [18][19][20]22,23 , jitter 20,23 , shimmer 20,23 , energy 18,19 , Zero-Crossing Rate (ZCR) 18,19 , and Mel-Frequency Cepstral Coefficients (MFCCs) 19 .…”