yongjoonlee/semambapp

SEMamba++ (Interspeech 2026 · Long Paper Track)

Official Hugging Face repository for SEMamba++. [Demo] [Paper (arXiv)] [Github]

SEMamba++ is a general speech restoration (GSR) framework that leverages global, local, and periodic spectral patterns via a Mamba-based architecture. It handles a range of degradation conditions including noise, reverberation, and clipping.

---

Prerequisites

Install all required dependencies:

pip install -r requirements.txt

For the Mamba backbone, follow the installation guide from SEMamba, which resolves CUDA-specific build issues.

---

Datasets

SEMamba++ can be trained on any dataset that provides speech, noise, and room impulse response (RIR) samples. Point each split to the corresponding JSON manifest file:

Split	File
Training speech	data/train_speech.json
Training noise	data/train_noise.json
Training RIR	data/train_rir.json
Validation (clean)	data/val_clean.json
Validation (degraded)	data/val_degraded.json

Download sources

• Speech: VCTK, LibriTTS
• Noise: DNS Challenge 2020, WHAM!
• RIR: Arni, DNS5

---

Pretrained weights

Pretrained weights will be released on HuggingFace.

The released model was trained on VCTK and LibriTTS (~500 hours of speech combined).

---

References

• SEMamba — Mamba-based speech enhancement backbone
• BigVGAN — Neural vocoder (NVIDIA)
• MPSENet — Multi-scale phase-aware speech enhancement

---

Citation

If you find SEMamba++ useful in your work, please cite:

@misc{lee2026semambageneralspeechrestoration,
  title         = {SEMamba++: A General Speech Restoration Framework
                   Leveraging Global, Local, and Periodic Spectral Patterns},
  author        = {Yongjoon Lee and Jung-Woo Choi},
  year          = {2026},
  eprint        = {2603.11669},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2603.11669}
}