Performance Improvement of Dysarthric Speech Recognition Using Context-Dependent Pronunciation Variation Modeling Based on Kullback-Leibler Distance

In this paper, we propose context-dependent pronunciation variation modeling based on the Kullback-Leibler (KL) distance for improving the performance of dysarthric automatic speech recognition (ASR). To this end, we construct a triphone confusion matrix based on KL distances between triphone models, and build a weighted finite state transducer (WFST) from the triphone confusion matrix. Then, dysarthric speech is recognized by a baseline ASR system. The corresponding phoneme sequence of the recognized sentence is then passed through the WFST to correct recognition errors. It is shown from dysarthric ASR experiments that average word error rate of an ASR system employing an error correction based on the proposed method is relatively reduced by 16.54% and 3.34%, compared to those of an ASR system without any error correction and using an error correction based on a conventional contextdependent phoneme confusion matrix, respectively.