Exploration of an Attention based Two-Stream Dual-Path Networks for Speech Enhancement

Deep neural networks are incapable of tackling inter-frame correlation information within speech sequences, which is important for speech enhancement task. Based on this consideration, attention mechanisms are designed to learn long-term temporal dependencies information from input features. This paper proposes an attention based Two-Stream Dual-Path Conformer Networks (TSDPCNet) for speech enhancement work, which can effectively extract both amplitude and phase information from speech signals. In this network, self-attention mechanism is employed to focus on the global information of the speech sequence, while the dual-path attention enables simultaneous extraction of sequence information from both the time and frequency domains. Experimental results on the VoiceBank + DEMAND datasets shows that the proposed TSDPCNet outperforms the conventional speech enhancement methods.