Dysarthric speech recognition based on error-correction in a weighted finite state transducer framework

Abstract

In this paper, a dysarthric speech recognition error-correction method in a weighted finite state transducer (WFST) framework is proposed to improve the performance of dysarthric automatic speech recognition (ASR). To this end, pronunciation variation models are constructed from a context-dependent confusion matrix based on a weighted Kullback-Leibler (KL) distance between triphones. Then, a WFST is finally constructed by combining the WFST of the baseline ASR, the constructed pronunciation variation models, a lexicon, and a language model. It is shown from the dysarthric ASR experiments that a WFST-based ASR system employing the proposed error-correction method achieves relative average word error rate reduction of 19.73%, compared to an ASR system without any error-correction method.