Objective assessment of autism spectrum disorder based on performance in structured interpersonal acting-out tasks with prosodic stability and variability.

In this study, we sought to objectively and quantitatively characterize the prosodic features of autism spectrum disorder (ASD) via the characteristics of prosody in a newly developed structured speech experiment. Male adults with high-functioning ASD and age/intelligence-matched men with typical development (TD) were asked to read 29 brief scripts aloud in response to preceding auditory stimuli. To investigate whether (1) highly structured acting-out tasks can uncover the prosodic of difference between those with ASD and TD, and (2) the prosodic stableness and flexibleness can be used for objective automatic assessment of ASD, we compared prosodic features such as fundamental frequency, intensity, and mora duration. The results indicate that individuals with ASD exhibit stable pitch registers or volume levels in some affective vocal-expression scenarios, such as those involving anger or sadness, compared with TD and those with TD. However, unstable prosody was observed in some timing control or emphasis tasks in the participants with ASD. Automatic classification of the ASD and TD groups using a support vector machine (SVM) with speech features exhibited an accuracy of 90.4%. A machine learning-based assessment of the degree of ASD core symptoms using support vector regression (SVR) also had good performance. These results may inform the development of a new easy-to-use assessment tool for ASD core symptoms using recorded audio signals.

Quantification of speech and synchrony in the conversation of adults with autism spectrum disorder.

Autism spectrum disorder (ASD) is a highly prevalent neurodevelopmental disorder characterized by impairments in social reciprocity and communication together with restricted interest and stereotyped behaviors. The Autism Diagnostic Observation Schedule (ADOS) is considered a 'gold standard' instrument for diagnosis of ASD and mainly depends on subjective assessments made by trained clinicians. To develop a quantitative and objective surrogate marker for ASD symptoms, we investigated speech features including F0, speech rate, speaking time, and turn-taking gaps, extracted from footage recorded during a semi-structured socially interactive situation from ADOS. We calculated not only the statistic values in a whole session of the ADOS activity but also conducted a block analysis, computing the statistical values of the prosodic features in each 8s sliding window. The block analysis identified whether participants changed volume or pitch according to the flow of the conversation. We also measured the synchrony between the participant and the ADOS administrator. Participants with high-functioning ASD showed significantly longer turn-taking gaps and a greater proportion of pause time, less variability and less synchronous changes in blockwise mean of intensity compared with those with typical development (TD) (p<0.05 corrected). In addition, the ASD group had significantly wider distribution than the TD group in the within-participant variability of blockwise mean of log F0 (p<0.05 corrected). The clinical diagnosis could be discriminated using the speech features with 89% accuracy. The features of turn-taking and pausing were significantly correlated with deficits of ASD in reciprocity (p<0.05 corrected). Additionally, regression analysis provided 1.35 of mean absolute error in the prediction of deficits in reciprocity, to which the synchrony of intensity especially contributed. The findings suggest that considering variance of speech features, interaction and synchrony with conversation partner are critical to characterize atypical features in the conversation of people with ASD.

Effect of learning on listening to ultra-fast synthesized speech.

A text-to-speech synthesizer that would produce easily understandable voices at very fast speaking rates is expected to help persons with visual disability to acquire information effectively with screen reading softwares. We investigated the intelligibility of Japanese text-to-speech systems at fast speaking rates, using four-digit random numbers as the vocabulary of the recall test. We also studied the fast and intelligible text-to-speech engine, using HMM-based synthesizer with the corpus with fast speaking rate. As the results, the statistical models trained with the fast speaking corpus was effective. The learning effect was significant in the early stage of the trials and the effect sustained for several weeks.