Time Domain Speech Enhancement With Attentive Multi-scale Approach

Speech enhancement aims to suppress the additive noise from noisy speech signals to improve the speech quality. It is believed that multi-scale temporal information learned from the speech inputs strengthens the mask prediction or noise sup-pression especially for the encoder-mask-decoder like structure in time-domain speech enhancement techniques. In this paper, we propose a multi-scale encoding and decoding scheme that captures multiple temporal resolutions for improving speech quality. We also propose an attention module to capture the global temporal information of each-scale embedding in the encoding layer. The experiments show that the proposed approach achieves 9.0% and 2.9% relative improvements over the best baseline in terms of perceptual evaluation of the speech quality (PESQ) and signal-to-distortion ratio (SDR), respectively.