Multi-Channel Speech Enhancement using a Minimum Variance Distortionless Response Beamformer based on Graph Convolutional Network

The Minimum Variance Distortionless Response (MVDR) beamforming algorithm is frequently utilized to extract speech and noise from noisy signals captured multiple microphones. A frequency-time mask should be employed compute the Power Spectral Density (PSD) matrices of signal interest obtain optimal weights for beamformer. Deep Neural Networks (DNNs) are widely used estimating time-frequency masks. This paper adopts a novel method using Graph Convolutional (GCNs) learn spatial correlations among different channels. GCNs integrated into embedding space U-Net architecture estimate Complex Ideal Ratio Mask (cIRM). We use cIRM in an MVDR beamformer further improve enhancement system. simulate room acoustics data experiment extensively with our approach types microphone array. Results indicate superiority when compared current state-of-the-art methods. metrics obtained by proposed significantly improved, except Scale-Invariant Source-to-Distortion (SI-SDR) score. Perceptual Evaluation Speech Quality (PESQ) score shows noticeable improvement over baseline models (i.e., 2.207 vs. 2.104 2.076). Our implementation can found following link: https://github.com/3i-hust-asr/gnn-mvdr-final.