β-Masking MMSE speech enhancement for speech recognition

Speech recognition performance deteriorates drastically when they are deployed in practical situation where the speech is corrupted by additive noise. One way to improve the robustness of the speech recognition system is to enhance the speech prior to its recognition. This paper focuses on developing a masking-based β-order minimum mean square error (β-masking MMSE) speech enhancement for speech recognition under noise condition. Addressing the artifacts introduced by enhancement algorithm and the remaining noise after denoising, we modified the estimation algorithm of spectral parameters for the β-masking MMSE by controlling the power of processing noise, strengthening the weak signal processing, oversuppressing the residual noise and reestimating a priori SNR. The evaluation shows the proposed enhancement scheme is significantly effective to improve the performance of state-of-the-art speech recognition.