A Multi-Microphone Speech Enhancement Algorithm Tested Using Acoustic Vector Sensors

Abstract

In this paper, we present a speech enhancement algo- rithm for multi-microphone systems that enhances a target signal in noisy multi-talker situations. We apply the general multichan- nel Wiener filtering framework, for which we have developed a new technique to directly estimate the auto-correlation of the target signal assuming its direction is known. The advantage of our approach compared to traditional multichannel Wiener filtering is that it effectively works with both non-stationary, speech-like noise interference and also stationary background noise. The estimation of the auto-correlation of the target signal is derived by minimizing the target signal power across pairs of microphones. For this reason, we refer to our algorithm as MWF- minPESP (Multichannel Wiener Filtering based on minimum Pair-wise Estimated Signal Power). In a sense, MWF-minPESP is a beamforming algorithm that minimizes the output signal power. We present an experiment that compares the performance of MWF-minPESP with traditional multichannel Wiener filtering and frequency-domain minimum variance distortionless response (FMV) beamforming using two acoustic vector sensors (AVS), comprising six microphone signals, in an office space. The results indicate that MWF-minPESP outperforms the other two state- of-the-art speech enhancement techniques.