Adaptation of a Pronunciation Dictionary for Dysarthric Speech Recognition

In the general framework of an automatic speech recognition system, a pronunciation dictionary, that is a mapping table from a phoneme sequence to a word, is used both in the processes of training and recognition. However, this pronunciation dictionary is not always adequate in the case of dysarthric speech recognition because dysarthric people often have difficulty pronouncing words in the same way they are pronounced in the dictionary. In this paper, we investigate the adaptation of a pronunciation dictionary to an individual dysarthric person and evaluate the effectiveness of adapting the dictionary using a dysarthric speech recognition task. In the proposed method, in order to find rules we can use to modify a dictionary, we analyze the pattern of mis-recognition in the phoneme recognition results. By following the extracted rules, we add pronunciations to the dictionary for the target dysarthric person. We evaluate the effectiveness of the adapted pronunciation dictionary on a continuous speech recognition task and demonstrate that the adapted dictionary can decrease the word error rate.