Multichannel MMSE Wiener Filter Using Complex Real and Imaginary Spectral Coefficients for Distributed Microphone Speech Enhancement

In this paper, the authors propose a frequency domain multichannel Wiener filter for distributed microphone speech enhancement using acoustic arrays. The current state-of-the-art single channel estimators achieve noticeable performance gains using the to-noise ratio (SNR) and segmental signal-to-noise ratio (SSNR) objective measures, which measure noise reduction, but only achieve marginal performance gains using the Log-Likelihood Ratio (LLR) and Perceptual Evaluation of Speech Quality (PESQ) objective metrics, which correlate better than SNR and SSNR with speech distortion and overall speech quality. By extending the traditional single channel Wiener filter to multiple distributed channels through minimum mean-square error (MMSE) estimation of the complex real and imaginary components, the approach presented here demonstrates increases in the SSNR, LLR, and PESQ objective measures. Experimental results show that the new multichannel Wiener filter using distributed microphones produces gains of 5.0 dB (SSNR improvement), 0.7 (LLR output), and 0.8 (PESQ output) averaged across the 0 dB, 5 dB, and 10 dB input SNRs over the baseline single channel Wiener filter.