Fusion of Multiple Audio Descriptors for the Recognition of Dysarthric Speech

Dysarthria is a motor speech disorder majorly caused from brain stroke or severe accidents affecting the left hemisphere of the brain. Due to the injury, speech muscles and tongue movement become weak, leading to highly unintelligible speech. In the past decade, the performance of Automatic Speech Recognition (ASR) systems have improved signif-icantly for normal speech, but when it comes to dysarthric speech, performance is far from satisfactory. In this pa-per we implement dysarthric speech recognition system in a speaker-adaptive manner, leveraging additional acoustic features to help the model effectively learn and adapt to each speaker's unique speech patterns. Feature selection is based on MFCC combined with different variants of Jitter and Shimmer. Despite the challenges posed by limited data and considerable variability among the speakers, our system achieves an average Word Recognition Accuracy (WRA) of 83.11 % on the UASpeech database, which is a significant improvement over previous studies. In addition to ASR, we also analyze dysarthric speakers with very low intelligibility rates and found some prominent insights into their speech characteristics.