Papers › MambAttention: Mamba with Multi-Head Attention for Generalizable Single-Channel Speech...

MambAttention: Mamba with Multi-Head Attention for Generalizable Single-Channel Speech Enhancement

1 Jul 2025arXiv:2507.00966archive 2025-07-28

Nikolai Lund Kühne, Jesper Jensen, Jan Østergaard, Zheng-Hua Tan

With the advent of new sequence models like Mamba and xLSTM, several studies have shown that these models match or outperform state-of-the-art models in single-channel speech enhancement, automatic speech recognition, and self-supervised audio representation learning. However, prior research has demonstrated that sequence models like LSTM and Mamba tend to overfit to the training set. To address this issue, previous works have shown that adding self-attention to LSTMs substantially improves generalization performance for single-channel speech enhancement. Nevertheless, neither the concept of hybrid Mamba and time-frequency attention models nor their generalization performance have been explored for speech enhancement. In this paper, we propose a novel hybrid architecture, MambAttention, which combines Mamba and shared time- and frequency-multi-head attention modules for generalizable single-channel speech enhancement. To train our model, we introduce VoiceBank+Demand Extended (VB-DemandEx), a dataset inspired by VoiceBank+Demand but with more challenging noise types and lower signal-to-noise ratios. Trained on VB-DemandEx, our proposed MambAttention model significantly outperforms existing state-of-the-art LSTM-, xLSTM-, Mamba-, and Conformer-based systems of similar complexity across all reported metrics on two out-of-domain datasets: DNS 2020 and EARS-WHAM_v2, while matching their performance on the in-domain dataset VB-DemandEx. Ablation studies highlight the role of weight sharing between the time- and frequency-multi-head attention modules for generalization performance. Finally, we explore integrating the shared time- and frequency-multi-head attention modules with LSTM and xLSTM, which yields a notable performance improvement on the out-of-domain datasets. However, our MambAttention model remains superior on both out-of-domain datasets across all reported evaluation metrics.

PaperPDFCode

Code

NikolaiKyhne/MambAttention officialmentioned in papermentioned on GitHubpytorch report
nikolaikyhne/xlstm-senet mentioned on GitHubpytorch report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Automatic Speech RecognitionMambaSpeech EnhancementSpeech Recognitionspeech-recognition

Datasets

Introduced by this paper, per the archive.

VB-DemandEx

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Speech Enhancement Deep Noise Suppression (DNS) Challenge MambAttention ESTOI 95.9 #4 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge MambAttention Number of parameters (M) 2.33 #4 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge MambAttention PESQ-WB 3.671 #4 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge MambAttention SI-SDR-WB 21.234 #4 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge MambAttention SSNR 15.116 #4 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge xLSTM-SENet ESTOI 95.4 #6 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge xLSTM-SENet Number of parameters (M) 2.20 #6 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge xLSTM-SENet PESQ-WB 3.588 #6 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge xLSTM-SENet SI-SDR-WB 20.854 #6 of 36 Archive leaderboard report
Speech Enhancement Deep Noise Suppression (DNS) Challenge xLSTM-SENet SSNR 14.526 #6 of 36 Archive leaderboard report
Speech Enhancement VB-DemandEx MambAttention ESTOI 0.801 #1 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MambAttention Number of parameters (M) 2.33 #1 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MambAttention PESQ (wb) 3.026 #1 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MambAttention SI-SDR 16.684 #1 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MambAttention SSNR 7.674 #1 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx SEMamba ESTOI 0.800 #2 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx SEMamba Number of parameters (M) 2.25 #2 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx SEMamba PESQ (wb) 3.002 #2 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx SEMamba SI-SDR 16.593 #2 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx SEMamba SSNR 7.590 #2 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx xLSTM-SENet ESTOI 0.795 #3 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx xLSTM-SENet Number of parameters (M) 2.20 #3 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx xLSTM-SENet PESQ (wb) 2.973 #3 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx xLSTM-SENet SI-SDR 16.414 #3 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx xLSTM-SENet SSNR 7.933 #3 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MP-SENet ESTOI 0.787 #4 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MP-SENet Number of parameters (M) 2.05 #4 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MP-SENet PESQ (wb) 2.935 #4 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MP-SENet SI-SDR 16.202 #4 of 4 Archive leaderboard report
Speech Enhancement VB-DemandEx MP-SENet SSNR 7.641 #4 of 4 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

LSTMMamba

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections