BUDDy: Single-channel Blind Unsupervised Dereverberation with Diffusion Models

This is the official code for our paper BUDDy: Single-channel Blind Unsupervised Dereverberation with Diffusion Models

We invite you to check our companion website for listening samples and insights into the paper

1 - Requirements

Install required Python packages with python -m pip install -r requirements.txt

2 - Checkpoints

You can access our pretrained checkpoint, trained on VCTK anechoic speech, at the following link

3 - Testing

You can launch blind dereverberation with bash test_blind_dereverberation.sh. You can launch informed dereverberation with bash test_informed_dereverberation.sh. In both cases, do not forget to add the path to the pretrained model checkpoint in the bash file (i.e. replace ckpt=<pretrained-vctk-checkpoint.pt> with your path). The directory tree in audio_examples/ contains an example test set to reproduce the results.

4 - Training

You can retrain an unsupervised diffusion model on your own dataset with bash train.sh. Do not forget to fill in the path to your training and testing dataset (i.e. replace dset.train.path=/your/path/to/anechoic/training/set and so on)

5 - Citing

If you used this repo for your own work, do not forget to cite us:

@bibtex

@article{moliner2024buddy,
    title={{BUDD}y: Single-channel Blind Unsupervised Dereverberation with Diffusion Models},
    author={Moliner, Eloi and Lemercier, Jean-Marie and Welker, Simon and Gerkmann, Timo and V\"alim\"aki, Vesa},
    year={2024},
    journal={arXiv 2405.04272}
}