The Impact of Speaker Diarization on DNN-based Autism Severity Estimation

Area of Science:

• Computational linguistics
• Developmental psychology
• Machine learning in healthcare

Background:

• Autism Spectrum Disorder (ASD) diagnosis relies on observational assessments.
• Objective, speech-based biomarkers for autism severity are needed.
• Automatic analysis of child speech presents unique challenges.

Purpose of the Study:

• To develop and evaluate a speech-based system for estimating autism severity.
• To compare two speaker diarization methods for identifying individual speech segments.
• To assess the system's performance using acoustic and prosodic speech features.

Main Methods:

• Speaker diarization using Mel-Frequency Cepstral Coefficients (MFCC) and pitch, or x-vectors (DNN embeddings).
• Training speaker diarization models with a Fully Connected Deep Neural Network (FCDNN).
• Developing a Convolutional Neural Network (CNN) for autism severity estimation using 48 speech features from 132 children recorded during ADOS assessments.

Main Results:

• The MFCC and Pitch method for speaker diarization yielded a better Diarization Error Rate (DER) of 26.91%.
• The autism severity estimation system achieved a Pearson correlation of 0.606 between predicted and actual ADOS scores.
• The system successfully identified children's speech segments and estimated autism severity.

Conclusions:

• Acoustic and prosodic speech features, combined with effective speaker diarization, can be used to estimate autism severity.
• The developed system demonstrates potential as a non-invasive tool for autism assessment.
• Further research can refine these speech-based methods for clinical application.