Combination of GMM-Based Speech Estimation Method and Temporal Domain SVD-Based Speech Enhancement for Noise Robust Speech Recognition

摘要

This study proposes a speech recognition method which is made robust to noise by combining speech signal estimation based on GMM and speech enhancement based on SVD in the temporal domain. Conventional speech signal estimation based on GMM has the problems that the time dependence of the noise is not considered and the performance is degraded in a low-SNR environment. As regards the first problem, successive updating of the mean noise vector is performed in this study to follow the time variation of the noise. As regards the second problem, an attempt is made to improve performance by improving the SNR beforehand by means of speech enhancement based on SVD in the time domain. Furthermore, in speech enhancement based on SVD in the time domain, the over-subtraction factor for the noise component is introduced in order to minimize the effect of noise, and adaptive determination of the factor is considered. The proposed method is evaluated using the AURORA2 database, and it is shown that the speech recognition accuracy is improved compared to conventional speech signal estimation based on GMM.