Multiframe Maximum Likelihood Distortionless Response Filter for Single-Channel Speech Enhancement

Speech enhancement is an essential and challenging task in the signal-processing field. In this article, a multiframe maximum likelihood distortionless response (MFMLDR) filter is proposed for single-channel speech enhancement. Specifically, in accordance with the multiframe signal model, the MFMLDR filter is obtained by constructing the log-likelihood objective function with a nondistorted speech correlation component constraint. Then, to optimize the performance of the MFMLDR filter, the interframe correlation matrix is estimated more accurately by leveraging the sparsity of speech in the time-frequency domain. Finally, an online MFMLDR (OL-MFMLDR) filter is presented by the recursive least squares (RLS) algorithm for online applications. The proposed method can make a good tradeoff between noise suppression and speech distortion reduction. Both objective and subjective evaluation results indicate that the proposed method outperforms the compared single-channel approaches in overall speech enhancement performance, particularly at low signal-to-noise ratios (SNRs).