Region Based Prediction and Score Combination for Automatic Intelligibility Assessment of Dysarthric Speech

Abstract

The intelligibility assessment of dysarthric speech is essential for planning therapy. Time-frequency representations have been used for automatic intelligibility assessment of dysarthria. These representations have been derived from the utterance as a whole. As voiced and unvoiced components have different characteristics; in this work, we use different time-frequency representations for voiced and unvoiced segments and use them for intelligibility assessment with CNN classifier. Finally, we combine the scores obtained by the two systems to assign an intelligibility level. The combined system's performance is found to be superior to the systems that used both voiced and unvoiced components separately or together as one utterance. The utterances of the TORGO database are used in the intelligibility assessment. Automatic assessment of speech intelligibility reduces speech-language pathologists' time and effort in assisting diagnosis and treatment design.