On the use of array learners towards Automatic Speech Recognition for dysarthria

Providing Automatic Speech Recognition (ASR) systems for dysarthria is a challenging task since the normal and the disabled speech have different attributes; hence, using ASR systems designed and trained for normal speakers is not an effective approach. It is important to craft ASR technologies specifically for the speech disabled. Nonetheless, because of the complexity and variability of dysarthric speech, previous studies failed to achieve adequate performance. In this paper we investigated the applications of array learners towards dysarthric speech recognition. The array was implemented by several neural networks that configured to work in parallel. The proposed approach was verified by using the speech materials of seven dysarthric subjects with speech intelligibility from 2% to 86%. For comparison, the results were compared with a dysarthric ASR based on the legacy single-learner approach as the reference model. It is shown that the array learner-based dysarthric ASR improved the mean word recognition rate of 10.41% over the reference model, and decreased the error rate of 4.84%.