Corpus Construction for Deaf Speakers and Analysis by Automatic Speech Recognition
Akio Kobayashi, Keiichi Yasu · 2023
This study explores automatic speech recognition (ASR) for the deaf and hard-of-hearing. Despite the recent progress in ASR for dysarthric speakers, existing research primarily focuses on people with motor speech disorders. Thus, the effect of speech diversity on the performance of ASR is not considered for ambiguous deaf speech owing to a lack of auditory feedback. Therefore, we compiled a corpus of speech of many profoundly deaf speakers to compare the ASR performance with that of normal-hearing speakers. The performance analysis is reported through a set of phoneme recognition experiments. Furthermore, we show that additional phonological features that reflect deaf speakers’ articulation can improve performance in phoneme recognition for deaf speech.