Benchmarking the Unsupervised Discovery of Phoneme Inventories With Discrete Speech Units

arXiv · GitHub · Website

DiscoPhon is a multilingual benchmark evaluating unsupervised phoneme discovery from discrete speech units. Given only 10 hours of speech in an unseen language, models must produce discrete units that map to a predefined phoneme inventory.

Getting started

DiscoPhon requires Python ≥ 3.12 and has no system dependencies.

• Install this package:

  pip install discophon            # core: data preparation and phoneme discovery
  pip install discophon[abx]       # adds ABX discriminability (fastabx)
  pip install discophon[baselines] # adds the baseline models

• Follow the tutorials to download data, evaluate models, and prepare your submission.
• Current leaderboard.

References

@misc{poli2026discophon,
  title={{DiscoPhon}: Benchmarking the Unsupervised Discovery of Phoneme Inventories With Discrete Speech Units},
  author={Maxime Poli and Manel Khentout and Angelo Ortiz Tandazo and Ewan Dunbar and Emmanuel Chemla and Emmanuel Dupoux},
  year={2026},
  eprint={2603.18612},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2603.18612},
}