A Novel Approach to Multi-Channel Speech Enhancement Based on Graph Neural Networks

Multi-channel speech enhancement aims at utilizing spatial relationships between signals captured from a microphone array along with temporal-spectral information efficiently to estimate the clean target. An emerging approach is to design deep learning-based end-to-end architectures. In this work, we provide a new way to process latent multi-channel representations. We introduce a novel end-to-end system called temporal graph convolutional network, which views the embedding space of multi-channel signals as a graph and leverages the power of graph neural networks (GNNs) to analyze spatial correlations as well as temporal-spectral information simultaneously. To be specific, graph convolutional networks (GCNs), a popular GNN variant, are integrated into a complex convolutional encoder-decoder structure to compute a complex ideal ratio mask. The estimated mask is subsequently multiplied with the reference microphone spectrogram to get enhanced speech. We demonstrate the superiority of our approach by comparing it to state-of-the-art methods on ConferencingSpeech 2021 Challenge data. Our results and analyses prove that GCN is a novel yet promising mechanism for speech enhancement systems, providing an interesting alternative for recent deep learning-based approaches and inspiration for future research.