coml/maubert-quantisers

MauBERT quantisers

The quantisation models used to produce the pseudo-labels of the MauBERT paper (ACL 2026). Each of them maps the frame-level features of one layer of coml/maubert-feat, coml/maubert-phone or the HuBERT base model to discrete units, and was fitted on 10 hours of one language.

The fine-tuned models themselves are not released: use maubert to extract the features of a pre-trained model, then one of these quantisers to transcribe them into units.

Layout

<model>/<dataset>/<language>/<n_clusters>cl.<ext>

manifest.tsv lists every file with its base model, layer, quantisation method, dataset, language and number of clusters.

model	base model	layer	method	file
fr-feat-simple	MauBERT-feat	feat_proj	feat-freq	.pt
fr-kmeans-feature	MauBERT-feat	feat_proj	kmeans	.km
fr-kmeans-hubert_9	MauBERT-feat	HuBERT layer 10	kmeans	.km
fr-phone-ft	MauBERT-feat	phone_proj	phone-ft	.ft
kmeans-hubert_11	HuBERT base	HuBERT layer 12	kmeans	.km
pr-kmeans-phone	MauBERT-phone	phone_proj	kmeans	.km
pr-kmeans-ups_proj	MauBERT-phone	ups_proj	kmeans	.km
pr-phone-ft	MauBERT-phone	phone_proj	phone-ft	.ft
pr-phone-simple	MauBERT-phone	phone_proj	phone-simple	.pt

The .km files are joblib pickles of a scikit-learn MiniBatchKMeans, the .ft files are torch state dicts of the adapted phone-prediction layer, and the other .pt files are torch tensors.

Datasets and languages

dataset	languages
discophon	cmn, deu, eng, eus, fra, jpn, swa, tam, tha, tur, ukr, wol
dev-languages	sw, ta, th, tr, uk
zrc2017 (ZeroSpeech 2017 test)	english, french, german (LANG1), mandarin, wolof (LANG2)
casual-languages	english, french

Every quantiser comes in a 100-cluster variant (370 for pr-phone-simple, which uses the full phone inventory of MauBERT-phone). The discophon languages additionally have a variant with as many clusters as the language has phones: cmn 43, deu 42, eng 40, eus 30, fra 35, jpn 43, swa 30, tam 30, tha 41, tur 28, ukr 36, wol 40.

Usage

Extract the features of the corresponding layer with maubert, then transcribe them with the quantiser:

hf download coml/maubert-quantisers fr-kmeans-hubert_9/discophon/deu/100cl.km --local-dir quantisers

python -m maubert.features MANIFEST_FILE FEATURE_DIR \
    --type maubert-feat --checkpoint coml/maubert-feat --layer 9

python -m maubert.transcribe FEATURE_DIR \
    quantisers/fr-kmeans-hubert_9/discophon/deu/100cl.km units.jsonl kmeans

Citing

@inproceedings{ortiztandazo-etal-2026-maubert,
    title = "{M}au{BERT}: Universal Phonetic Inductive Biases for Few-Shot Acoustic Units Discovery",
    author = "Ortiz Tandazo, Angelo  and
      Khentout, Manel  and
      Benchekroun, Youssef  and
      Hueber, Thomas  and
      Dupoux, Emmanuel",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.24/",
    doi = "10.18653/v1/2026.acl-long.24",
    pages = "568--585",
    ISBN = "979-8-89176-390-6",
}