TEA-PSE 2.0: Sub-Band Network for Real-Time Personalized Speech Enhancement
Yukai Ju, Shimin Zhang, Wei Rao, Yannan Wang, Tao Yu, Lei Xie, Shidong Shang
Abstract
Personalized speech enhancement (PSE) utilizes additional cues like speaker embeddings to remove background noise and interfering speech and extract the speech from target speaker. Previous work, the Tencent-Ethereal-Audio-Lab personalized speech enhancement (TEA-PSE) system, ranked 1st in the ICASSP 2022 deep noise suppression (DNS2022) challenge. In this paper, we expand TEA-PSE to its sub-band version - TEA-PSE 2.0, to reduce computational complexity as well as further improve performance. Specifically, we adopt finite impulse response filter banks and spectrum splitting to reduce computational complexity. We introduce a time frequency convolution module (TFCM) to the system for increasing the receptive field with small convolution kernels. Besides, we explore several training strategies to optimize the two-stage network and investigate various loss functions in the PSE task. TEA-PSE 2.0 significantly outperforms TEA-PSE in both speech enhancement performance and computation complexity. Experimental results on the DNS2022 blind test set show that TEA-PSE 2.0 brings 0.102 OVRL personalized DNSMOS improvement with only 21.9% multiply-accumulate operations compared with the previous TEA-PSE.