Single-channel speech dereverberation based on block-wise weighted prediction error and nonnegative matrix factorization

Abstract

This paper proposes a speech dereverberation method based on a block-wise weighted prediction error (BWPE) method and nonnegative matrix factorization (NMF). The proposed BWPE method iteratively estimates late reverberation using maximum likelihood (ML) estimation in a block-wise manner. To ensure consistent de-reverberation performance over time, a forgetting factor is applied on intermediate estimates. Thus, the recent statistics of the signal are emphasized during the block-wise processing. In addition, the NMF-based source separation method is applied to reduce early reverberation that remains in the signal processed by the proposed BWPE method. The performance of the proposed method is compared with that of the conventional weighted prediction error (WPE) method by measuring the Segmental signal-To-noise ratio (SSNR). It is shown from the comparison that the proposed method achieves a higher SSNR than the conventional method. Moreover, the proposed method can be implemented in a real-Time audio recording device with an algorithmic delay of 20ms.