Dysarthric Speech Recognition Using Pseudo-Labeling, Self-Supervised Feature Learning, and a Joint Multi-Task Learning Approach

In this paper, we investigate the use of the spontaneous speech of dysarthric people for training an automatic speech recognition (ASR) model for them. Although the spontaneous speech of dysarthric people can be collected relatively easily compared to script-reading speech, which is obtained by having them read a prepared script, labeling the spontaneous speech of dysarthric people is very difficult and costly. For training an ASR model using unlabeled speech data, pseudo-labeling and self-supervised feature learning have been studied as effective approaches; however, the effectiveness of these approaches has not been clear when they are applied to the unlabeled dysarthric speech. In addition, pseudo-labeling may not be effective since the pseudo-labels of dysarthric speech include many errors and are not reliable. In this paper, we evaluate the above two approaches for the dysarthric speech recognition, and we propose a multi-task learning approach, which combines these approaches to train an ASR model that is robust against the errors in the pseudo-labels. Experimental results using Japanese and English datasets demonstrated that all approaches are effective, but among them, the proposed multi-task learning approach showed the best performance.