A Speech Distortion Weighted Single-Channel Wiener Filter Based STFT-Domain Noise Reduction

In this work, we focus on the single-channel noise reduction (NR) in the short-time Fourier transform (STFT) domain from the traditional signal processing perspective. As conventional single-channel NR methods suffer from a serious speech distortion (SD), we propose an SD weighted single-channel Wiener filter (SDW-SWF), where an auxiliary parameter µ is exploited to trade-off the SD and residual noise variance. In the subspace, the obtained SDW-SWF can be formulated as a function of µ and a set of generalized eigenvectors of correlation matrices. In addition, we theoretically analyze the impacts of the trade-off factor and the rank on the SD, residual noise power and the output signal-to-noise ratio (SNR). Finally, numerical results validate the effectiveness of the proposed method, exhibiting a consistency with the theoretical findings. It can be concluded that the SDW-SWF approach enables more degrees-of-freedom to improve the speech intelligibility at a sacrifice of SNR.