DBHN-Net: Dual-Branch Hybrid Neural Network for Low-Complexity Monaural Speech Enhancement

Area of Science:

• Artificial Intelligence
• Signal Processing
• Computer Science

Background:

• Artificial neural networks (ANNs) offer high performance for speech enhancement (SE) but suffer from high computational and energy demands.
• Spiking neural networks (SNNs) show promise for reduced power consumption but often incur information loss due to discrete activations and complex dynamics.

Purpose of the Study:

• To develop a novel speech enhancement method that balances performance with reduced computational complexity and energy consumption.
• To address the information loss challenge in SNNs while leveraging their power efficiency.

Main Methods:

• Proposed a Dual-Branch Hybrid Neural (DBHN) Network integrating ANN and SNN branches.
• Introduced BandSplit and Time-Frequency (TF)-Mamba modules for energy compression and performance enhancement.
• Utilized Spiking Feature Extraction Group (SFEG) and Information Transformation Block (ITB) with residual connections to minimize information loss.
• Designed Interaction and TF-Cross Attention-Fusion modules for effective inter-branch information exchange and fusion.

Main Results:

• The DBHN model achieved superior performance across three public datasets.
• Demonstrated an average 7.5-fold reduction in computational complexity compared to baseline models.
• Successfully mitigated information loss issues common in SNNs.

Conclusions:

• The proposed DBHN network offers an effective solution for low-power, high-performance speech enhancement.
• The hybrid architecture successfully combines the strengths of ANNs and SNNs.
• This approach facilitates the practical deployment of advanced speech enhancement in resource-constrained environments.