Research map / Architectures and efficiency / Attention and transformers
BSDB-Net: Band-Split Dual-Branch Network with Selective State Spaces Mechanism for Monaural Speech Enhancement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BSDB-Net, a band-split dual-branch (magnitude and complex-spectrum) speech enhancement network using Mamba selective state-space blocks for time and frequency modeling, cutting compute roughly 8x versus baselines while staying competitive.
Problem
Complex-spectrum speech enhancement suffers from a compensation effect between amplitude and phase, and state-of-the-art SE models have high computational complexity that limits deployment in real-time or resource-constrained settings.
Approach
Decouple magnitude and complex (RI) spectra into parallel branches (MEN and CEN) with interaction modules, compress the frequency axis via band-splitting, and replace attention/RNN sequence modeling with bidirectional Mamba along frequency and unidirectional Mamba along time.
Key results
- On WSJ0-SI84 Set-A, BSDB-Net reports average PESQ 2.87, ESTOI 77.04%, SI-SDR 11.05 dB, slightly above GaGNet on PESQ and ESTOI.
- On VoiceBank+Demand, BSDBNet(256) reports PESQ-WB 3.11 and COVL 3.73, outperforming listed baselines such as FDFNet (3.05 PESQ).
- Replacing Mamba with Transformer raises MACs from 1.68 G/s to 28.06 G/s with similar PESQ (2.90 vs 2.92); the paper reports about 8.3x average complexity reduction versus baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use parallel magnitude and complex-spectrum branches instead of either branch alone (ablation) | BSDB-Net · WSJ0-SI84 | PESQ: 2.67 / 2.76 -> 2.92 | Helped |
| use parallel magnitude and complex-spectrum branches instead of either branch alone (ablation) | BSDB-Net · WSJ0-SI84 | ESTOI: 71.25 / 74.32 -> 78.19 | Helped |
| use parallel magnitude and complex-spectrum branches instead of either branch alone (ablation) | BSDB-Net · WSJ0-SI84 | SI-SDR: 9.84 / 10.37 -> 11.34 | Helped |
| increase Mamba width and depth from 64-4 to 128-6 (ablation) | BSDB-Net · WSJ0-SI84 | PESQ: 2.62 -> 2.92 | Helped |
| replace Mamba sequence modeling with LSTM (ablation) | BSDB-Net · WSJ0-SI84 | PESQ: 2.92 -> 2.82 | Hurt |
| replace Mamba sequence modeling with Transformer (ablation) | BSDB-Net · WSJ0-SI84 | PESQ: 2.92 -> 2.90 | Hurt |
| replace Mamba sequence modeling with Transformer (ablation) | BSDB-Net · WSJ0-SI84 | MACs: 28.06G/s -> 1.68G/s | Hurt |
| use BSDB-Net instead of GaGNet | speech enhancement · Set-A unseen-speaker test | PESQ Avg.: 2.81 -> 2.87 | Helped |
| use BSDB-Net instead of GaGNet | speech enhancement · Set-B unseen-speaker test | SI-SDR Avg.: 13.93 -> 14.39 | Helped |
| use BSDBNet(256) instead of FDFNet | speech enhancement · VoiceBank+Demand | PESQ-WB: 3.05 -> 3.11 | Helped |
| use BSDBNet(256) instead of FDFNet | speech enhancement · VoiceBank+Demand | COVL: 3.65 -> 3.73 | Helped |
Surprising findings
- The Transformer variant reaches nearly the same PESQ as the Mamba variant (2.90 vs 2.92), so the main gain is complexity rather than quality.
- DBTNet, a non-causal model with about 42.64 G/s MACs, slightly outperforms the proposed causal model on PESQ.
Limitations
- The model is causal, which the authors cite as a reason for slightly lower performance than non-causal DBTNet.
- Mamba is described as still immature for deployment, requiring further optimization.
- Performance gain over the Transformer variant is relatively minor.
Experimental setup
Datasets: WSJ0-SI84 DNS-Challenge noise VoiceBank+Demand
Benchmarks: WSJ0-SI84 (Set-A, Set-B) VoiceBank+Demand
Metrics: PESQ ESTOI SI-SDR STOI CSIG CBAK COVL MACs parameters
Baselines: ConvTasNet DPRNN DDAEC LSTM CRN GCRN DCCRN FullSubNet CTSNet GaGNet DBTNet SEGAN
Builds on: Mamba FullSubNet CTSNet GaGNet
Closest papers by meaning
- Complex-Cycle-Consistent Diffusion Model for Monaural Speech Enhancement (AAAI 2025 · similarity 0.88)
- TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation (ICLR 2025 · similarity 0.82)
- Rethinking Training Targets, Architectures and Data Quality for Universal Speech Enhancement (NeurIPS 2026 · similarity 0.79)
- Shallow Diffusion for Fast Speech Enhancement (Student Abstract) (AAAI 2024 · similarity 0.79)
- RealMAN: A Real-Recorded and Annotated Microphone Array Dataset for Dynamic Speech Enhancement and Localization (NeurIPS 2024 · similarity 0.79)
- Audio Super-Resolution with Latent Bridge Models (NeurIPS 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.