Linear prediction filtering on cepstral time series for noise-robust speech recognition

In this paper, we propose adopting the algorithm of linear prediction coding (LPC) to proceeds the temporal feature streams in speech recognition for noise robustness. Using LPC, an FIR filter can be obtained and applied to the time series of Mel-frequency cepstral coefficients (MFCC), and in general the fast-varying component in the modulation spectrum of MFCC can be alleviated accordingly. We have found that the smoothing of MFCC modulation spectrum helps to reduce the noise effect and enhance noise robustness of MFCC. Experiments conducted on the Aurora-2 connected digit database shows that the proposed LPC-wise method improves the recognition accuracy of MVN- and HEQ-preprocessed MFCC under a wide range of noise-corrupted situations.