Subspace-Based Speech Correlation Vector Estimation for Single-Microphone Multi-Frame MVDR Filtering

Abstract

Aiming at exploiting the speech correlation across consecutive timeframes in the short-time Fourier transform domain, the multi-frame minimum variance distortionless response (MFMVDR) filter for single-microphone speech enhancement has been proposed. This filter is designed to avoid speech distortion while minimizing the total signal output power. To compute the MFMVDR filter, an estimate of the highly time-varying normalized speech correlation vector is required. In this paper, we propose a subspace-based estimator for the normalized speech correlation vector based on the Q largest eigenvalues and their corresponding eigenvectors of the prewhitened noisy speech correlation matrix. Experimental results for different speech signals, noise types and signal-to-noise ratios show that the proposed subspace-based estimator yields the best results in terms of speech quality and noise reduction compared to a state-of-the-art maximum-likelihood estimator.