Kai Li1, Guo Chen1, Xiaolin Hu1
1Tsinghua University, China
ArXiv | Demo
# SPMamba: State-space model is all you need in speech separation ## Abstract SPMamba is an innovative speech separation model designed to address the complexity of modeling long audio sequences in existing LSTM and Transformer-based systems. Building on the robust TF-GridNet architecture, SPMamba replaces traditional BLSTM components with bidirectional Mamba modules, which efficiently capture spatiotemporal relationships in both time and frequency dimensions. This allows the model to handle long-range dependencies with linear computational complexity. By leveraging bidirectional processing, SPMamba enhances separation performance by utilizing both past and future context. Extensive experiments on datasets such as **WSJ0-2Mix, WHAM!, Libri2Mix**, and the newly constructed **Echo2Mix** demonstrated that SPMamba not only outperformed state-of-the-art models but also reduced computational complexity. ## 🔥 News [2024-11-22] Release Checkpoint SPMamba model checkpoints (Libri2Mix and Echo2Mix) are available at [[Checkpoint]](https://github.com/JusperLee/SPMamba/releases/tag/v1.0) [2024-09-06] Demo Website SPMamba is now available at [[Demo]](https://cslikai.cn/SPMamba/) [2024-09-06] Release Datasets Echo2Mix, a new dataset for speech separation. [[DataEcho2Mix]](https://drive.google.com/file/d/1nJ9ujAbf4LxXEFzFwEpr9CwJOeNHghw0/view) [2024-05-09] Update SPMamba **WHAM!** Result: SI-SNRi=17.4 dB, SDRi=17.6 dB [2024-04-23] Update SPMamba MACs: **238.21 G/s** using [[code]](https://github.com/state-spaces/mamba/issues/110) [2024-04-18] Update SPMamba **WSJ0-2Mix** Result: SI-SNRi=22.5 dB, SDRi=22.7 dB ## Installation clone the repository ```bash git clone https://github.com/JusperLee/SPMamba.git && cd SPMamba conda env create -f look2hear.yml conda activate look2hear ``` ## Usage To train the SPMamba model, run the following command: ```bash python audio_train.py --conf_dir=configs/spmamba.yml ``` ## Performance *Here, you can include a brief overview of the performance metrics or results that SPMamba achieves using WSJ0-2Mix, WHAM!, Libri2Mix, Echo2Mix*  ## License SPMamba is licensed under the Apache License 2.0. For more details, see the [LICENSE](LICENSE) file in the repository. ## Acknowledgements SPMamba is developed by the **Look2Hear** at Tsinghua University. We would like to thank the **ESPnet team** for their contributions to the open-source community and for providing a solid foundation for our work. ## Citation If you use SPMamba in your research or project, please cite the following paper: ``` @article{li2024spmamba, title={SPMamba: State-space model is all you need in speech separation}, author={Li, Kai and Chen, Guo and Hu, Xiaolin}, journal={arXiv preprint arXiv:2404.02063}, year={2024} } ``` ## Contact For any questions or feedback regarding SPMamba, feel free to reach out to us via email: `tsinghua.kaili@gmail.com`