Multi-stream gated and pyramidal temporal convolutional neural networks for audio-visual speech separation in multi-talker environments

Speech separation is the task of extracting target speech from noisy mixture. In applications like video telephones or video conferencing, lip movements of the target speaker are accessible, which can be leveraged for speech separation. This paper proposes a time-domain audio-visual speech separation model under multi-talker environments. The model receives audio-visual inputs including noisy mixture and speaker lip embedding, and reconstructs clean speech waveform for the target speaker. Once trained, the model can be flexibly applied to unknown number of total speakers. This paper introduces and investigates the multi-stream gating mechanism and pyramidal convolution in temporal convolutional neural networks for audio-visual speech separation task. Speaker- and noise-independent multi-talker separation experiments are conducted on GRID benchmark dataset. The experimental results demonstrate the proposed method achieves 3.9 dB and 1.0 dB SI-SNRi improvement when compared with audio-only and audio-visual baselines respectively, showing effectiveness of the proposed method.