Robust speech recognition for similar Japanese pronunciation phrases under noisy conditions

This paper proposes a new noisy robust speech recognition method. Under noise circumstances, several noise reduction methods have been developed and they are applied in various noise conditions. However, in case of similar pronunciation speech, for example, it is still not easy to realize high recognition accuracy. In this paper, the new processing algorithm into speech modulation spectrum is proposed as running spectrum analysis (RSA) and it is adequately applied to observed speech data. Using this method, the proposed system can improve about 1 – 4 % compared to current conventional methods.