Blind Extraction-Based Multichannel Speech Enhancement in Noisy and Reverberation Environments

Abstract

Speech enhancement has important applications in sensor, hearing aids, robotics, and video conferencing. However, the speech enhancement performance is severely deteriorated by additional background noise and high reverberations. To solve the problem of speech enhancement in noisy and acoustically reverberant scenarios, this letter proposes a multichannel speech enhancement algorithm based on blind extraction to achieve speech denoising and dereverberation. First, a new model for speech enhancement is constructed by assuming the reverberations generated by later reflections as additional and unrelated noise components. Subsequently, a blind signal extraction approach is designed to extract the direct sound and early reflected sounds, achieving dereverberation and denoising. Experimental results confirm that the proposed algorithm achieves better speech enhancement in noisy and acoustic reverberation scenarios and that the effect of dereverberation and noise reduction is superior to that of popular speech enhancement algorithms, especially in high reverberation environments.