A new robust LPCCS for speech recogniton in channel distortion environments

Linear prediction (LP) analysis plays an important role in speech processing. linear prediction cepstral coefficients (LPCCs) are proved to be better features for speech recognition. In LP analysis, the generation of speech is modeled with an autoregressive (AR) random process. In fact, speech from vocal cord goes through vocal tract first, and then goes through channel. However, general LP analysis does not consider the effect of channel and the LPCCs show little robustness in speech recognition under channel distortion. In this paper, distorted speech is modeled by two different AR models which denote vocal tract and channel, respectively. After the poles are computed, K-means is used to cluster all poles and a new method is proposed to choose Channel poles. For each frame of test utterances, instead of designing a filter using channel poles, robust LPCCs are computed directly by removing channel poles from all poles. Experiment shows that this approach can obtain a better performance than general methods.