Keyword spotting is necessary for triggering human-machine speech interaction. It is a challenging task especially in low signal-to-noise ratio and moving scenarios, such as on a sweeping robot with strong ego-noise. This paper proposes a novel approach for joint ego-noise suppression and keyword detection. The keyword detection model accepts outputs from multi-look adaptive beamformers. The noise covariance matrix in the beamformer is in turn updated using the keyword absence probability given by the model, forming an end-to-end loop-back. The keyword model also adopts a multi-channel feature fusion using self-attention, and a hidden Markov model for online decoding. The performance of the proposed approach is verified on real-word datasets recorded on a sweeping robot.
0 commit comments