Direct modulation on LPC coefficients with application to speech enhancement and improving the performance of speech recognition in noise

A novel method of noise reduction of speech based on direct modulation of LPC (linear predictive coding) coefficients is proposed. This method introduces higher-order derivatives of LPC coefficients with respect to the noise-to-signal energy ratio (NSR). With these derivatives, the noisy LPC coefficients are refined flexibly and efficiently to reduce noise contaminations. This method only needs the environmental NSR, and does not require knowledge of the probability distribution of the noise. This enhancement method is incorporated in an HMM (hidden Markov model)-based speech recognition system using LPC-derived cepstral features. A pronounced recognition error rate reduction is obtained after the speech enhancement.<>