DPCCN: Densely-Connected Pyramid Complex Convolutional Network for Robust Speech Separation and Extraction

In recent years, a number of time-domain speech separation methodshave been proposed. However, most of them are very sensitiveto the environments and wide domain coverage tasks. In thispaper, from the time-frequency domain perspective, we propose adensely-connected pyramid complex convolutional network, termedDPCCN, to improve the robustness of speech separation under complicatedconditions. Furthermore, we generalize the DPCCN to targetspeech extraction (TSE) by integrating a new specially designedspeaker encoder. Moreover, we also investigate the robustness ofDPCCN to unsupervised cross-domain TSE tasks. A Mixture-Remixapproach is proposed to adapt the target domain acoustic characteristicsfor fine-tuning the source model. We evaluate the proposedmethods not only under noisy and reverberant in-domain condition,but also in clean but cross-domain conditions. Results show that forboth speech separation and extraction, the DPCCN-based systemsachieve significantly better performance and robustness than the currentlydominating time-domain methods, especially for the crossdomaintasks. Particularly, we find that the Mixture-Remix finetuningwith DPCCN significantly outperforms the TD-SpeakerBeamfor unsupervised cross-domain TSE, with around 3.5 dB SISNR improvementon target domain test set, without any source domain performancedegradation.