Temporal Modulation Spectral Restoration for Robust Speech Recognition

In this paper, we propose a temporal modulation spectral resto-ration (TMSR) approach for robust feature extraction in automatic speech recognition. There were three main function blocks in TMSR. First, mean and variance normalization (CMVN) was applied to the original feature sequence. Second, the noise characteristic was estimated with an analysis of the normalized features. Third, a gain function was designed to attenuate noise and enhance speech components from the normalized features. In this study, a simple high-pass filter noise estimation scheme and a gain function derived by the generalized maximum a posteriori ( GMAP) algorithm were employed in TMSR. The proposed method was evaluated on two benchmark databases, Aurora-3 and Aurora-4. Results showed that TMSR outperformed the baseline and several well-known robust feature extraction methods.