Speech Dereverberation Based on Sparse Matrix Decomposition

Because of the increasingly demands of high quality audio signal, speech dereverberation, as the preliminary processing of speaker recognition and automatic speech recognition(ASR), becomes more and more important. The speech obtained from microphones is always distorted by reverberation. Conventional approaches always build a model to dereverberate speech. However, in different environments, these models may not be effective. For this reason, we propose an algorithm which does not base on any environment model assumptions so that it can be used for all speech. A piece of clean speech can be represented through a sparse matrix. The reverberated speech matrix can be decomposed into two matrices, clean speech matrix and reverberated noise matrix, to capture the sparse components of the speech using Robust Principal Component Analysis (RPCA). Evaluations via many different criterions show that the new approach preserves the clean speech's information well and dereverberate the speech well.