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.
CITATION STYLE
Dong, J., Zhou, D., & Zhang, Q. (2015). Robust Feature Extraction Based on Teager-Entropy and Half Power Spectrum Estimation for Speech Recognition. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 9426, pp. 91–101). Springer Verlag. https://doi.org/10.1007/978-3-319-26181-2_9
Mendeley helps you to discover research relevant for your work.