wav2vec 2.0 / HuBERT

weights

open(PyTorch, TensorFlow and safetensors checkpoints on the Hub, ungated, plus fairseq .pt files)

data

open(LibriSpeech (CC BY 4.0) and Libri-Light, both public

code

open(fairseq pretraining and CTC fine-tuning recipes for wav2vec 2.0 and HuBERT)

license

Apache-2.0(the wav2vec 2.0 and HuBERT checkpoints on the Hub

research-release

voxpopuli(2021 VoxPopuli checkpoints under CC-BY-NC-4.0, published with a separate dataset paper

The wav2vec 2.0 and HuBERT checkpoints are Apache-2.0, they were trained on public LibriSpeech and Libri-Light audio, and fairseq publishes the pretraining and fine-tuning recipes, so the models can be rebuilt from what is released. Meta's later VoxPopuli checkpoints carry a non-commercial license but belong to a separate research release.

• https://huggingface.co/api/models?author=facebook&search=wav2vec2&limit=1000&expand[]=downloads&expand[]=createdAt&expand[]=tags&expand[]=gated recorded 2026-09-27

  Family listing: the wav2vec2 base, large and XLS-R checkpoints tagged "license:apache-2.0"; the *-voxpopuli* checkpoints tagged "license:cc-by-nc-4.0".

• https://huggingface.co/api/models/facebook/wav2vec2-base-960h?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified&expand[]=tags&expand[]=siblings recorded 2026-09-27

  Hub record for wav2vec2-base-960h: "gated":false, tag "license:apache-2.0", and model.safetensors in the file list.

• https://huggingface.co/facebook/wav2vec2-base-960h/raw/main/README.md recorded 2026-09-27

  Card: "The base model pretrained and fine-tuned on 960 hours of Librispeech on 16kHz sampled speech audio."

• https://raw.githubusercontent.com/facebookresearch/fairseq/main/examples/hubert/README.md recorded 2026-09-27

  HuBERT README: "HuBERT Large (~316M params) | [Libri-Light](https://github.com/facebookresearch/libri-light) 60k hr"; "## Train a new model".

• https://raw.githubusercontent.com/facebookresearch/fairseq/main/examples/wav2vec/README.md recorded 2026-09-27

  wav2vec README: "### Train a wav2vec 2.0 base model"; "### Fine-tune a pre-trained model with CTC".

• https://raw.githubusercontent.com/facebookresearch/fairseq/main/README.md recorded 2026-09-27

  fairseq README: "fairseq(-py) is MIT-licensed." and "The license applies to the pre-trained models as well."

• https://www.openslr.org/12/ recorded 2026-09-27

  OpenSLR LibriSpeech page: "<b>License:</b> CC BY 4.0".

Hugging Face downloads of the five declared wav2vec 2.0 and HuBERT checkpoints; the base encoder is pulled mostly as a starting point for fine-tuning.

• https://huggingface.co/api/models?author=facebook&search=hubert&limit=100&expand[]=downloads&expand[]=createdAt&expand[]=tags&expand[]=gated recorded 2026-09-27

  Trailing-30-day downloads: hubert-large-ls960-ft 313,174, hubert-base-ls960 243,099.

• https://huggingface.co/api/models?author=facebook&search=wav2vec2&limit=1000&expand[]=downloads&expand[]=createdAt&expand[]=tags&expand[]=gated recorded 2026-09-27

  Trailing-30-day downloads: wav2vec2-base 2,161,483, wav2vec2-base-960h 1,421,400, wav2vec2-large-960h-lv60-self 152,118.

An English recognizer from 2020 whose strength today is as a pretrained encoder to fine-tune. Its reported results cover LibriSpeech only, and it has no place on the public boards that rank Whisper, so it sits a step below Whisper.

• https://artificialanalysis.ai/speech-to-text recorded 2026-09-27

  The speech-to-text table lists no row for this model.

• https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/d2c5b384deccdb82834f41aeaffcc618c00efa2f/english_short_latest.csv recorded 2026-09-27

  The results CSV lists no facebook/wav2vec2 or facebook/hubert row among its 71 models.

• https://huggingface.co/facebook/wav2vec2-base-960h/raw/main/README.md recorded 2026-09-27

  Card results table: "| 3.4 | 8.6 |".

• https://huggingface.co/facebook/wav2vec2-large-960h-lv60-self/raw/main/README.md recorded 2026-09-27

  Card results table under "clean" and "other": "| 1.9 | 3.9 |".