BMG-Net: Gated Speech Enhancement Network Integrating Bi-Directional Mamba and Masked Separable Attention
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Transformer-based speech enhancement models often suffer from quadratic computational complexity ($O\left(L^{2}\right)$) and struggle to accurately disentangle speech in low-SNR environments. To address these challenges, we propose BMG-Net, which features a novel Bi-Mamba-MSA Hybrid (BM-MH) bottleneck. Adopting a “divide-and-conquer” strategy, this module comprises a Global Expert using Bi-directional Mamba for linearcomplexity long-range modeling, and a Local Expert utilizing Masked Separable Attention (MSA) to separate speech from noise via a foreground/background head strategy. Furthermore, a pixelwise Attention Gate (AG) mechanism is designed to dynamically fuse global semantics with local details. Extensive experiments on VoiceBank+DEMAND demonstrate that BMG-Net outperforms SOTA models (e.g., MP-SENet, CMGAN) in PESQ and CSIG while maintaining lower computational overhead.
Fields of Study & Subjects
Speech and Audio Processing
Speech Recognition and Synthesis
Image Enhancement Techniques
Publication details
- Journal / source
- International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
- Pages
- 1258–1262
- Publication date
- 16 January 2026
- Language
- EN