Skip to content

addse

Code for "Absorbing Discrete Diffusion for Speech Enhancement".

Documentation

Repository

Audio examples

Installation

  • Install uv

  • Clone the repository and install dependencies:

    git clone [email protected]:philgzl/addse.git && cd addse && uv sync

Data preparation

Training data

The following datasets are used:

Place each dataset under data/external/. Then run the following scripts:

./ldopt_bigspeech.sh
./ldopt_bignoise.sh

This converts the data to an optimized format for litdata and writes it in data/chunks/.

Alternatively, update the two shell scripts to use your own speech and noise data.

Validation data

Validation data is directly streamed from Hugging Face. No need to prepare anything.

Alternatively, update the configuration files in configs/ to use your own litdata-optimized validation data.

Evaluation data

Download the Clarity speech dataset to data/external/Clarity/. Then run:

uv run addse ldopt data/external/Clarity/ data/chunks/clarity/ --num-workers 4

The remaining evaluation data is directly streamed from Hugging Face.

Alternatively, update the configuration files in configs/ to use your own litdata-optimized evaluation data.

Training

To train a model:

uv run addse train configs/<model_name>.yaml

Checkpoints and metrics are written to logs/<model_name>/.

You can use the --wandb option to log metrics to W&B, and the --log_model option to additionally upload checkpoints to W&B, after configuring a .env with your credentials.

Evaluation

To evaluate a trained model:

uv run addse eval configs/<model_name>.yaml logs/<model_name>/checkpoints/last.ckpt --num-consumers 4

The results are written in eval.db by default.

Trained checkpoints

Will be released soon.

Citation

@article{gonzalez2026absorbing,
  title={Absorbing Discrete Diffusion for Speech Enhancement},
  author={Gonzalez, Philippe},
  journal={arXiv preprint arXiv:2602.22417},
  year={2026}
}