Koel Labs - Machine Learning

Contains the EDA, training, evaluation, and data processing code for Koel Labs. Evaluation results will be made available via Hugging Face Leaderboards. Cleaned datasets and model weights will also be made available via Hugging Face.

Read about all our repositories here.

Setup

Checkout the guides directory for standalone guides on finetuning, evaluation, dataset processing, and other topics. These can be run independently of the setup for the rest of the codebase, e.g., in a Colab notebook.

See the DEVELOPMENT.md for alternative setup instructions and details.

0. git clone https://github.com/KoelLabs/ML.git
1. Install Python 3.10.16
2. Duplicate the .env.example file and rename it to .env. Fill in the necessary environment variables.
3. Run the commands in './scripts/install.sh', e.g., with . ./scripts/install.sh.

Contributing

Check out the CONTRIBUTING.md for specific guidelines on contributing to this repository.

License

The code in this repository is licensed under the GNU Affero General Public License.

With the exception of a few models and Huggingface spaces released during the builders program under the Mozilla Public License, all Huggingface models and code will be released under the GNU Affero General Public License.

We retain all rights to the Koel Labs brand, logos, blog posts and website content.

Citation

If you use any of our models, datasets, or code for a publication, please cite our paper:

@inproceedings{metzger2026scaling,
  title={Scaling Human and G2P Supervision for Robust Phonetic Transcription},
  author={Metzger, Alexander and Srivastava, Aruna and Mukhamedvaleev, Ruslan},
  booktitle={Proceedings of Interspeech},
  year={2026},
  eprint={2606.16019},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}