Monaural Speech Enhancement Using a Multi-Branch Temporal Convolutional Network

Deep learning has achieved substantial improvement on single-channel speech enhancement tasks. However, the performance of multi-layer perceptions (MLPs)-based methods is limited by ability to capture long-term effective history information. The recurrent neural networks (RNNs), e.g., long short-term memory (LSTM) model, are able temporal dependencies, but come with issues high latency and complexity training.To address these issues, convolutional network (TCN) was proposed replace RNNs in various sequence modeling In this paper we propose a novel TCN model that employs multi-branch structure, called (MB-TCN), for monaural enhancement.The MB-TCN exploits split-transform-aggregate design, which expected obtain strong representational power at low computational complexity.Inspired TCN, incorporates one dimensional causal dilated CNN residual expand receptive fields capturing contextual information.Our extensive experimental investigation suggests MB-TCNs outperform (ResLSTMs), (TCNs), employ dense aggregations terms intelligibility quality, while providing superior parameter efficiency. Furthermore, our results demonstrate multiple state-of-the-art deep learning-based five widely used objective metrics.