A unified approach of compensation and soft masking incorporating a statistical model into the Wiener filter

In this paper, we present a new single-channel noise reduction method that integrates compensation and soft masking into the same statistical model assumptions for noise-robust speech recognition. By utilizing a Gaussian mixture model(GMM) as a pre-knowledge of speech and added noise signals, the proposed method can effectively restore clean speech spectra and separate out ambient noises from a target speech in the Wiener filter framework. The soft mask methods originally attempted to separate out the speech signal of the speaker of interest from a mixture of speech signals. In the proposed method, by using pre-trained speech and noise models, the soft mask techniques can be applied to separate out added noises from the target speech. Combined with the model-based Wiener filter performing compensation on the power spectrum, the technique can efficiently reduce distortions caused by nonstationary noises and finally reconstruct clean speech spectra from noise-corrupted observation. By applying the result in order to infer the a priori SNR of the Wiener filter, we can estimate the clean speech signal with greater accuracy. While the conventional Wiener filter causes inevitable distortions owing to noise reduction and does not solve non-stationary noises overlapped with speech presence periods, the proposed method can considerably solve these problems through compensation and softmasking based on speech and noise GMMs. The results evaluated in a practical speech recognition system for car environments show that the proposed method outperforms conventional methods.