In this paper, we present a novel architecture for multi-channel speech enhancement using a cross-channel attention-based Wave-U-Net structure. Despite the advantages of utilizing spatial information as well as spectral information, it is challenging to effectively train a multi-channel deep learning system in an end-to-end framework.
With a channel-independent encoding architecture for spectral estimation and a strategy to extract spatial information through an inter-channel attention mechanism, we implement a multi-channel speech enhancement system that has high performance even in reverberant and extremely noisy environments.
Experimental results show that the proposed architecture has superior performance in terms of signal-to-distortion ratio improvement (SDRi), short-time objective intelligence (STOI), and phoneme error rate (PER) for speech recognition.