Automatic creation of a Vowel Dataset for performing Prosody Analysis in ASD screening

Autism Spectrum Disorder (ASD) is a term used to describe a constellation of early-onset social communication deficits and repetitive sensorimotor behaviours associated with a strong genetic component as well as other causes. This paper aims at creating a tool for automatically isolating segments of the speech useful for extract prosody features for identifying children with ASD. In particular, in this first phase of the research, we are interested in the creation of a large dataset of 'a' vowels of ASD and not ASD people. The 'a' vowel contains relevant information on the voice quality and emotional states. The proposed methodology is divided into 2 phases. In the former the input audio is analyzed to determine the vowel onset and offset points, useful to extract the vowel regions. Then a spectrogram graphically visualizing the identified vowels is provided as input to the second phase, where a convolutional neural network classifies whether the histogram represents the vowel 'a'. The convolutional network reaches an average accuracy of 95.00% (standard deviation ± 2.60%) on a dataset of 640 samples with Stratified 5-Fold Cross-Validation.