Densely Connected Multi-Stage Model With Channel Wise Subband Feature For Real-Time Speech Enhancement

Research on single channel speech enhancement (SE) has a long tradition, but two main practical problems still remain unsolved. First, high quality enhancement, computational efficiency and low-latency are hard to be satisfied simultaneously in the existing practical systems. Second, specific scenario enhancement, such as singing and emotional speech, is also a intricate problem for conventional methods. In this paper, we propose a computationally efficient real-time speech enhancement network with densely connected multi-stage structures, which progressively enhances the channels-wise subband speech. The enhancing speech from earlier stage is used to guide the processing of deeper stage in oder to obtain coarse to fine estimates. Besides, the supervision is applied to all intermediate results that stabilizes the training and accelerates the convergence. Moreover, a adaptive fine-tune step is utilized with some small datasets of specific scenarios, which achieves superb improvement under corresponding scenes. As a result, the proposed method achieves promising performance in terms of speech quality and demonstrates robustness in complex scenarios.