New robust speech recognition using DTW in noise

This paper proposes a new robust speech recognition method. Since the hidden Markov model (HMM) algorithm need a lot of training calculation, The dynamic time warping (DTW) algorithm based on median filter is used instead in our system. According to the short-term energy method, the non-speech segment can be removed. Recognition accuracy is thus improved. The cepstral mean subtraction (CMS), running spectrum filtering (RSF) and dynamic range adjustment (DRA) methods are also applied for noise reduction. Our results show that the RSF/DRA and CMS/DRA are good. The median filter improved the accuracy of DTW and The recognition accuracy almost are same and that one using HMM in the 10 dB and 20 dB white noise.