runngezhang / causal-se

Official Implementation of "Inference and Denoise: Causal Inference-based Neural Speech Enhancement"

Inference and Denoise: Causal Inference-based Neural Speech Enhancement

This is the official implementation of "Inference and Denoise: Causal Inference-based Neural Speech Enhancement".
We are updating our code and additional supplementary material and soon. Thank you.

Requirements

• pytorch 1.12.1
• pytorch_lightning 1.7.5
• SOTA SE models (slightly modified to fit the usage of CISE) are credited to:
  • MANNER
  • CMGAN
  • DEMUCS

Usage

To train and inference, please run

python main.py \
  --save_path <path/to/checkpoint/and/tensorboard/> \
  --output_dir <path/to/prediction> \
  --use_pretrained \
  --batch_size 8 \
  --hop_length 320

References

@article{hsieh2022inference,
  title={Inference and Denoise: Causal Inference-based Neural Speech Enhancement},
  author={Hsieh, Tsun-An and Yang, Chao-Han Huck and Chen, Pin-Yu and Siniscalchi, Sabato Marco and Tsao, Yu},
  journal={arXiv preprint arXiv:2211.01189},
  year={2022}
}

Acknowledgments

• Bio-ASP Lab, CITI, Academia Sinica, Taipei, Taiwan

• Georgia Tech

• NTNU / U Kore

• IBM Research AI