Far Field Speech Enhancement at Low SNR in Presence of Nonstationary Noise Based on Spectral Masking and MVDR Beamforming

Abstract

Low Signal to Noise Ratio (SNR) conditions are highly likely during remote speech acquisition. This paper handles a method of remote speech multi-channel signal processing for speech enhancement in presence of strong nonstationary noise. The presented approach builds upon the Minimum Variance Distortionless response (MVDR) method, additionally filtering the multi-channel signal prior to MVDR beamforming coefficient estimation with a spectral mask. This mask is obtained by applying mixture observation vector clustering based on a spatial correlation model, which is estimated by a Complex Gaussian Mixture Model (CGMM). The posterior probabilities obtained during the CGMM Expectation-Maximization (EM) algorithm are used to estimate the cumulative noise mask, which is applied to the mixture. The masked mixture is then used to calculate the MVDR covariance matrix and beamforming coefficients. The method is tested on four mixtures acquired using a 66 microphone array at various low SNR. The results are compared to conventional MVDR and several other methods and validated using the Signal to Distortion Ratio (SDR) improvement metric. The results show that the presented method gives SDR improvement no less than 1–1.5 dB in the majority of cases, compared to MVDR, and performs best specifically at low SNR of \(-15\) – \(-20\) dB.