Benchmarking the Unsupervised Discovery of Phoneme Inventories With Discrete Speech Units

arXiv · DOI · 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.

• Install this package:

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

  Only the baselines extra has a system dependency: FFmpeg, required by torchcodec to read audio.
• Follow the tutorials to download data, evaluate models, and prepare your submission.
• Current leaderboard.

References

@inproceedings{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},
  booktitle = {{Interspeech 2026}},
  pages     = {6664--6669},
  doi       = {10.21437/Interspeech.2026-2791},
  issn      = {2958-1796},
}