Sepformer Dns4 16k Enhancement

🚀 SepFormer trained on Microsoft DNS-4 (Deep Noise Suppression Challenge 4 – ICASSP 2022) for speech enhancement (16k sampling frequency)

This repository offers all the essential tools for speech enhancement (denoising) using a SepFormer model implemented with SpeechBrain. The model is trained on 1300HRS of the Microsoft-DNS 4 dataset at a 16k sampling frequency. For a better experience, we recommend learning more about SpeechBrain. The evaluation results on the DNS4 2022 baseline dev set using DNSMOS are as follows:

✨ Features

• Multilingual Support: Supports languages such as English, German, Russian, French, Italian, and Spanish.
• Audio-to-Audio Task: Specialized in audio-to-audio tasks, particularly speech enhancement.
• Trained on Specific Datasets: Trained on the Microsoft DNS-4 dataset for deep noise suppression.
• Multiple Metrics Evaluation: Evaluated using metrics like SI-SNR, PESQ, SIG, BAK, and OVRL.

📦 Installation

First of all, please install SpeechBrain with the following command:

pip install speechbrain

Please notice that we encourage you to read our tutorials and learn more about SpeechBrain.

💻 Usage Examples

Basic Usage

from speechbrain.inference.separation import SepformerSeparation as separator
import torchaudio

model = separator.from_hparams(source="speechbrain/sepformer-dns4-16k-enhancement", savedir='pretrained_models/sepformer-dns4-16k-enhancement')

# for custom file, change path
est_sources = model.separate_file(path='speechbrain/sepformer-dns4-16k-enhancement/example_dns4-16k.wav') 

torchaudio.save("enhanced_dns4-16k.wav", est_sources[:, :, 0].detach().cpu(), 16000)

Advanced Usage

To perform inference on the GPU, add run_opts={"device":"cuda"} when calling the from_hparams method.

📚 Documentation

Model Evaluation Results

DNSMOS - deep noise suppression (DNS)- mean opinion score (MOS) is a non - intrusive evaluation metric. It computes 3 scores– SIG (speech quality), BAK (background noise quality), and OVRL (overall quality) each on a scale of 1 to 5, with 5 being the best quality.

Model Index

Limitations

The SpeechBrain team does not provide any warranty on the performance achieved by this model when used on other datasets.

🔧 Technical Details

The model is a SepFormer implemented with SpeechBrain and trained on 1300HRS of the Microsoft - DNS 4 dataset with a 16k sampling frequency.

📄 License

This project is licensed under the "apache - 2.0" license.

Referencing

Referencing SpeechBrain

@misc{speechbrain,
  title={{SpeechBrain}: A General-Purpose Speech Toolkit},
  author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio},
  year={2021},
  eprint={2106.04624},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  note={arXiv:2106.04624}
}

Referencing SepFormer

@inproceedings{subakan2021attention,
      title={Attention is All You Need in Speech Separation}, 
      author={Cem Subakan and Mirco Ravanelli and Samuele Cornell and Mirko Bronzi and Jianyuan Zhong},
      year={2021},
      booktitle={ICASSP 2021}
}

Referencing ICASSP 2022 Deep Noise Suppression Challenge

@inproceedings{dubey2022icassp,
  title={ICASSP 2022 Deep Noise Suppression Challenge},
  author={Dubey, Harishchandra and Gopal, Vishak and Cutler, Ross and Matusevych, Sergiy and Braun, Sebastian and Eskimez, Emre Sefik and Thakker, Manthan and Yoshioka, Takuya and Gamper, Hannes and Aichner, Robert},
  booktitle={ICASSP},
  year={2022}
}

About SpeechBrain

• Website: https://speechbrain.github.io/
• Code: https://github.com/speechbrain/speechbrain/
• HuggingFace: https://huggingface.co/speechbrain/