Speech separation has been very successful with deep learning techniques. In this paper, we propose Time-Frequency (T-F) domain path scanning network (TFPSNet) for speech separation task. By introduces T-F scan path to the network, frequency bins in different frames and frequency can interact directly. We also introduce a T-F path loss to improve the performance further. The proposed TFPSNet could learn more details of frequency structure and separate the feature in T-F domain. Experiments on public WSJ0-2mix datasets show that our approach outperforms the current state-of-the-arts (SOTA) method, and achieves 21.1dB SISDR improvement (SISDRi). Our approach has good generalization ability. The model trained on WSJ0-2mix dataset achieves 18.8dB SISDRi on Libri-2mix test set without any fine tuning work.

https://ieeexplore.ieee.org/abstract/document/9747554