Reference-Based Speech Enhancement via Feature Alignment and Fusion Network

Introduction

This repo provides a PyTorch implementation of the paper: Reference-Based Speech Enhancement via Feature Alignment and Fusion Network.

Requirements

pytorch>=1.10.0

torchaudio>=0.10.0

numpy>=1.21.0

tqdm>=4.64.0

pesq>=0.0.3

Usage

1. Replace the vbd dataset path in dataset.py and ref_prepare.py .
2. python ref_prepare.py to generate 15s reference clips.
3. python test_stage_1.py to get the result of one stage model.
4. python test_stage_2.py to get the result of two stage model.

Citation

If you find the code useful in your research, please cite:

@inproceedings{yue2022reference,
  title={Reference-based speech enhancement via feature alignment and fusion network},
  author={Yue, Huanjing and Duo, Wenxin and Peng, Xiulian and Yang, Jingyu},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={10},
  pages={11648--11656},
  year={2022}
}

LICENSE

MIT-LICENSE