Noise-robust linear prediction cepstral features for network speech recognition

Abstract

In this paper, we propose a perceptually-motivated method for modifying the speech power spectrum to obtain a set of linear prediction coding (LPC) parameters that possess good noiserobustness properties in network speech recognition. Speech recognition experiments were performed to compare the accuracy obtained from MFCC features extracted from AMR-coded speech that use these modified LPC parameters, as well as from LPCCs extracted from AMR bitstream parameters. The results show that when using the proposed LP analysis method, the recognition performance was on average 1.2% - 6.1% better than when using the conventional LP method, depending on the recognition task.