Diffusion Speech Quality Assesment (SQA)

This repository contains the official PyTorch implementations for the paper:

• Danilo de Oliveira, Julius Richter, Jean-Marie Lemercier, Simon Welker, Timo Gerkmann, "Non-intrusive Speech Quality Assessment with Diffusion Models Trained on Clean Speech", accepted at ISCA Interspecch 2025. [bibtex]

The code is largely based on the repository of [1]

Installation

• Create a new virtual environment with Python >= 3.11 (we have not tested other Python versions, but they may work).
• Install the package dependencies via pip install -r requirements.txt and let pip resolve the dependencies for you
• Install the EDM2 code as a submodule: git submodule update --init --recursive

Data

For training, we use the EARS-WHAM dataset.

Pretrained checkpoints

• The checkpoint used in the paper can be downloaded here

Training

Training is done by executing train_sqa.py. A minimal running example with default settings (as in our paper) can be run with

torchrun --standalone --nproc_per_node=<num-gpus> train_sqa.py --outdir=<log-dir> --data=<path-to-trainset> --batch-gpu=<batch-size-per-gpu>

where <path-to-trainset> should be a path to a folder containing clean .wav files (subdirectories are also supported).

EMA Reconstruction

To reconstruct a new EMA profile with length 0.08, run

python edm2/reconstruct_phema.py --indir=<log-dir> --outdir=<reconstructed-ema-dir> --outstd=0.080

For more detailed on post-hoc EMA reconstruction, please refer to the EDM2 repository.

SQA

To calculate the diffusion log likelihoods on a test set and save them in a csv file, run

torchrun --standalone --nproc_per_node=<num-gpus> calculate_likelihood.py --checkpoint=<path-to-pkl> --data_dir=<path-to-testset> --output_file=<path-to-csv>

The --checkpoint parameter should be the path to a snapshot or a reconstructed EMA profile.

License

The code and checkpoints are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Citations / References

We kindly ask you to cite our papers in your publication when using any of our research or code:

@inproceedings{deoliveira2025nonintrusive,
  title={Non-intrusive Speech Quality Assessment with Diffusion Models Trained on Clean Speech},
  author={{de Oliveira}, Danilo and Richter, Julius and Lemercier, Jean-Marie and Welker, Simon and Gerkmann, Timo},
  booktitle={Proc. Interspeech 2025},
  pages={2330--2334},
  year={2025}
}

[1] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, Samuli Laine, "Analyzing and Improving the Training Dynamics of Diffusion Models", CVPR 2024. [Code]