coml/hubert-base-vp20

HuBERT VP-20

HuBERT VP-20 is a HuBERT base model pretrained on a subset of 6k hours and 20 languages of VoxPopuli (all EU languages except English, French, and German) for the DiscoPhon benchmark. It was pretrained using the minimal_hubert library.

You can load it with HuggingFace Transformers:

from transformers import HubertModel

model = HubertModel.from_pretrained("coml/hubert-base-vp20")

Or with minimal_hubert:

from minimal_hubert import HuBERT, HuBERTPretrain

# Standard model
model = HuBERT.from_pretrained("coml/hubert-base-vp20")
# With pretraining head for classification
model_for_pretraining = HuBERTPretrain.from_pretrained("https://huggingface.co/coml/hubert-base-vp20/resolve/main/it2.pt")

Check out minimal_hubert if you are interested in pretraining or want to load HuBERT checkpoints from different libraries.

Files:

• model.safetensors and config.json: HuggingFace Transformers checkpoint and config.
• it1.pt: 1st iteration checkpoint.
• it2.pt: 2nd iteration checkpoint. Converted to HuggingFace state_dict to get model.safetensors.
• km100-mfcc.joblib: K-means trained on MFCCs of VoxPopuli-20. Used to train the 1st iteration.
• km500-it1-l10.joblib: K-means trained on features from the 10th layer of the 1st iteration model. Used to train the 2nd iteration.
• km256-it2-l11.joblib: K-means trained on features from the 11th layer of the 2nd iteration model. Used for DiscoPhon finetuning.

Citing

@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},
}