Neural sidelobe canceller for target speech separation

Abstract

A system configured to perform neural sidelobe cancelling to achieve enhanced target speech separation. A device improves target speech separation by combining a fixed directional beamformer with a deep neural network (DNN) configured to (i) selectively enhance speech in one target direction, (ii) cancel speech from non-target directions, and (iii) cancel non-speech interference from all directions. By combining the directional output from the beamformer with raw microphone signals as input signals to the DNN, the DNN is given a directional cue with which to select the target speech. The DNN generates mask data representing the target speech using a combination of temporal convolutional blocks and causal self-attention blocks. For example, the DNN achieves low latency by processing a large number of short frames, while the causal self-attention blocks smooth the output and reinforce local consistency of speech.