Japanese Speaker-Independent Homonyms Speech Recognition

Japanese has homonyms such as “hashi” ((Chop-sticks)) and “hashi” ((Bridge)). Word speech recognition has been studied for a long time, but homonym speech recognition in Japanese has not been studied. In this paper, we studied speaker-independent homonym speech recognition. For homonym speech recognition, pitch extraction has been normally used to estimate a pitch frequency.. However, we did not use pitch extraction in our study. Instead, we used an accent model that was a phoneme label with more length, Mora position, accent type and accent high or low. It means that we used the effect of pitch on formant. The results of the experiments were that 89% accuracy was obtained by using MFCC, full covariance HMM, and the accent model.