Exploring Phone Recognition in Pre-verbal and Dysarthric Speech

Abstract

In this study, we perform phone recognition on speech utterances made by two groups of people: adults who have speech articulation disorders and young children learning to speak language. We explore how these utterances compare against those of adult English-speakers who don’t have speech disorders, training and testing several HMM-based phone-recognizers across various datasets. Experiments were carried out via the HTK Toolkit with the use of data from three publicly available datasets: the TIMIT corpus, the TalkBank CHILDES database and the Torgo corpus. Several discoveries were made towards identifying best-practices for phone recognition on the two subject groups, involving the use of optimized Vocal Tract Length Normalization (VTLN) configurations, phone-set reconfiguration criteria, specific configurations of extracted MFCC speech data and specific arrangements of HMM states and Gaussian mixture models.