Robust Feature Extraction Based on Teager-Entropy and Half Power Spectrum Estimation for Speech Recognition

Abstract

In this paper, we present a robust feature extraction scheme for speech recognition. Compared to standard mel-frequency cepstral coefficients (MFCC), it incorporates perceptual information into half parameter spectrum not into the whole classical spectrum, and combines with Teager-Entropy to construct a new feature vector. Its performance is compared with several techniques, and detailed comparative performance analysis with various types of noise and a wide range of SNR values is presented. The results suggest that our feature achieves superior robustness with HMM-based recognizer on an English digit task. The 8.87 % reduction of average error rate is obtained in comparison to ordinary MFCC. Furthermore, the results also uncover that the half power spectrum-based method leads to superior performance over the whole power spectrum-based method in most given environment.