Speech Enhancement for Noise and Acoustic Reverberation Scenarios

‍ ‍The purpose of speech enhancement is to extract a pure target speech signal from a noisy speech signal. However， the sound source signal received in a reverberation environment is a collection of the target source signal and many delayed and attenuated reflections， which significantly reduces the quality and intelligibility of the target speech. To explore the problem of speech enhancement in noisy and acoustic reverberation scenarios， this study proposes an unsupervised multichannel speech enhancement method based on blind signal extraction when the prior information of the target speech and acoustic environment is unknown. First， a new speech enhancement model with noise and acoustic reverberation is constructed by considering the reverberation generated by later reflections as additional and unrelated noise components， and the target signal is implicitly modeled through a time-frequency mask using the primal-dual splitting algorithm. Subsequently， the cepstrum threshold method is used to enhance the harmonic structure of the target speech signal， enhancing the target speech signal in the noisy reverberation speech signal and attenuating other components with less energy than the target speech signal. Finally， as the interference signal on each channel is attenuated， the extracted target speech signal in each iteration has better exclusivity and is unmixed， an adaptive time-frequency Wiener masking inverse filtering is designed to enhance dereverberation and denoising. An experiment was conducted to evaluate and analyze the performance of dereverberation and denoising for actual speech signals under noisy and reverberation conditions. The experimental results demonstrated that the proposed algorithm has excellent performance in dereverberation and denoising. Additionally， we verified that the enhancement effect of the proposed algorithm is superior to several popular multi-channel speech enhancement algorithms.