Comparison of syllable-based and phoneme-based DNN-HMM in Japanese Speech Recognition

Japanese is syllabic language. Additionally we have studied syllable-based GMM-HMM for Japanese speech recognition. In this paper, we investigate the differences of recognition accuracy using phoneme/syllable-based GMM-HMM and DNN (Deep Neural Network)-HMM. First, we present a comparison of syllable-based and phoneme-based DNN-HMM. Second, we train the tied state left-context dependent syllable DNN-HMM, and compare these three types of modeling method. In the experiment, we obtained a 5% relative gain for WER using left-context syllable DNN-HMM in comparison with a left-context syllable GMM-HMM, and an 11% relative gain for WER using triphone DNN-HMM in comparison with a syllable-based DNN-HMM. Finally, we got results that modeling left-context phoneme has not worked and context independent syllable-based DNN-HMM got the best performance in the experiments, when applied to the ASJ+JNAS corpus, which consists of about 70 hours.