README.md

metadata

license: mit
pretty_name: DiscoPhon artifacts
tags:
  - speech
  - phoneme-discovery
  - discrete-units
  - discophon

DiscoPhon artifacts

Discrete units and evaluation scores of the models on the DiscoPhon leaderboard.

Baselines

Key	Model
spidr-mmsulab	coml/spidr-mmsulab
spidr-vp20	coml/spidr-vp20
hubert-mmsulab-it2	coml/hubert-base-mmsulab
hubert-vp20-it2	coml/hubert-base-vp20

The finetuned checkpoints are in coml/discophon-finetuned-baselines, under the same names.

Layout

One directory per model, then one directory per evaluation run:

{key}/
├── info.json
└── {condition}/{folder}/{layer}/
    ├── units-{language}-{split}.jsonl
    └── scores.jsonl

• condition: zero-shot, or ft-{language}-{duration} for the model finetuned on duration (10min, 1h, 10h) of training data in language.
• folder: how the representations of the layer are evaluated:
  • many_to_one: 256 units, each mapped to its most frequent phoneme.
  • one_to_one: as many units as phonemes, plus one, mapped by a bijection. The directory also holds the K-means models (km{K}-*.joblib, with K the number of units).
  • continuous: ABX on the continuous representations. Scores only, no units.
• layer: Transformer layer, starting at 1.

info.json gives the step between units (in ms) and the layers reported on the leaderboard, for each track and condition ("0" for zero-shot, "10h" for finetuned on 10h):

{"step_units": 20, "layers": {"many_to_one": {"0": 6, "10h": 6}, "one_to_one": {"0": 6, "10h": 6}}}

The other runs (other layers, finetuning durations, and languages) are provided for analysis.

What is included

For the four baselines:

Folder	Conditions	Layers	Languages evaluated
many_to_one	zero-shot, finetuned on 10min, 1h, 10h	SpidR: 5 to 12, HuBERT: 1 to 12	all 12
continuous	zero-shot, finetuned on 10min, 1h, 10h	1 to 12	all 12
one_to_one	zero-shot, finetuned on 10h	leaderboard layer	all 12 (zero-shot), finetuning language (finetuned)

SpidR units come from the prediction heads of the layer. HuBERT units come from a K-means on the layer, trained on the pretraining data (zero-shot) or the finetuning data (finetuned).

Languages: cmn, deu, eng, eus, fra, jpn, swa, tam, tha, tur, ukr, wol. Splits: dev, test.

File formats

Units, one utterance per line, as read by the discophon evaluation:

{"file": "F07_O_001", "units": [75, 131, 159, 69, 215, 215, ...]}

Scores, one score per line, as written by python -m discophon.benchmark:

{"language": "deu", "split": "test", "metric": "per", "score": 0.6574108529121169}

Metrics: per, r_val, f1, pnmi (phoneme discovery), triphone_abx_discrete_{within,across}_speaker (ABX on units), and triphone_abx_continuous_{within,across}_speaker (ABX on continuous representations).

Usage

Download the scores of one model, or the units of one run:

hf download coml/discophon-artifacts --repo-type dataset --include "spidr-vp20/*/scores.jsonl" --local-dir artifacts
hf download coml/discophon-artifacts --repo-type dataset --include "spidr-vp20/zero-shot/many_to_one/6/*" --local-dir artifacts

Evaluate the units of a run again with discophon:

python -m discophon.benchmark /path/to/discophon_data artifacts/spidr-vp20/zero-shot/many_to_one/6 scores.jsonl --kind many-to-one

Export the leaderboard scores of a model:

python -m discophon.leaderboard export artifacts/spidr-vp20

Submissions

To add a model to the leaderboard, open a pull request on this dataset with a directory following the layout above, and one on the GitHub repository with the leaderboard scores. See the submission guide.

Citation

@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},
}