Optimization of dysarthric speech recognition

Abstract

Explores the residual vocal ability of people who have severe motor impairments accompanied with severe dysarthria, and develops methods for improving the performance of automatic speech recognition (ASR) of dysarthric speech. The target applications for this technology are in the development of communication and control devices for these people. In our speech recognition system, we developed an adaptive word detection algorithm to detect words in highly irregular dysarthric speech. We also implemented perceptually-based mel frequency cepstrum coefficients (MFCC) for the parametric representation of the speech signal, and we adopted the left-to-right discrete hidden Markov model (DHMM) for speech pattern recognition. The system was tested with one person who has cerebral palsy and dysarthria, reducing the intelligibility of her speech to less than 15%. Our initial results on a word set consisting of ten digits demonstrated that recognition rates above 90% can be achieved if more than ten repetitions are used for training.