| --- |
| license: mit |
| tags: |
| - wav2vec2 |
| - forced-alignment |
| - torchaudio |
| - automatic-speech-recognition |
| --- |
| |
| # wav2vec2 base 960h — torchaudio checkpoint, rehosted unmodified |
|
|
| The checkpoint behind `torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H`, byte-for-byte as PyTorch |
| publishes it. **Nothing here is modified, retrained, quantized or converted** — this is a rehost, not |
| a derivative. |
|
|
| | | | |
| |---|---| |
| | File | `wav2vec2_fairseq_base_ls960_asr_ls960.pth` | |
| | Size | 377,664,473 bytes | |
| | SHA-256 | `488fd4f16de84438ffc945334278c1b9fb9b7159a806c1080b16111a958c945d` | |
| | Copied from | <https://download.pytorch.org/torchaudio/models/wav2vec2_fairseq_base_ls960_asr_ls960.pth> | |
| | Model | wav2vec 2.0 BASE, pre-trained and fine-tuned for ASR on 960 h of LibriSpeech | |
|
|
| Rehosted for [Thematic Thing](https://thematicthing.com), an on-device qualitative-analysis |
| application, which uses it for CTC forced alignment (word-level timestamps) and needs the artefact it |
| downloads to come from a source it controls, with a written licence chain. |
|
|
| ## Usage |
|
|
| `torchaudio` resolves this file by name from its own CDN. To use this copy instead, place it in the |
| torch hub checkpoint cache under its original filename before building the bundle — `torch.hub` |
| short-circuits on an existing file and makes no HTTP call: |
|
|
| ```python |
| import torch, torchaudio |
| dst = Path(torch.hub.get_dir()) / "checkpoints" / "wav2vec2_fairseq_base_ls960_asr_ls960.pth" |
| # ... download this repo's file to `dst`, verify its SHA-256, then: |
| model = torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H.get_model() |
| ``` |
|
|
| ## Licence and attribution |
|
|
| The weights are the work of **Meta Platforms (Facebook, Inc. and its affiliates)**, published with |
| the *wav2vec 2.0* release in [`facebookresearch/fairseq`](https://github.com/facebookresearch/fairseq) |
| under the **MIT License**, and redistributed by PyTorch under that same licence. torchaudio states |
| this in its own source, in the docstring of the bundle that names this exact file |
| ([`torchaudio/pipelines/_wav2vec2/impl.py`](https://github.com/pytorch/audio/blob/v2.8.0/src/torchaudio/pipelines/_wav2vec2/impl.py), |
| v2.8.0, lines 343–354): |
|
|
| > Originally published by the authors of *wav2vec 2.0* under MIT License and redistributed with the |
| > same license. |
| > [[License](https://github.com/facebookresearch/fairseq/blob/ce6c9eeae163ac04b79539c78e74f292f29eaa18/LICENSE), |
| > [Source](https://github.com/facebookresearch/fairseq/blob/ce6c9eeae163ac04b79539c78e74f292f29eaa18/examples/wav2vec#pre-trained-models)] |
|
|
| That pinned `LICENSE` is the MIT License, **Copyright (c) Facebook, Inc. and its affiliates**, and is |
| included here verbatim as `LICENSE-fairseq-MIT.txt`. PyTorch is doing exactly what this repository |
| does — hosting a copy of the fairseq weights under the licence they were released with. |
|
|
| The `.pth` container is torchaudio's own parameter layout for the same weights, so torchaudio's |
| **BSD 2-Clause** licence (Copyright (c) 2017 Facebook Inc.) is included as |
| `LICENSE-torchaudio-BSD-2-Clause.txt`. |
|
|
| Please cite the authors of the model: |
|
|
| ```bibtex |
| @article{baevski2020wav2vec, |
| title={wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations}, |
| author={Baevski, Alexei and Zhou, Henry and Mohamed, Abdelrahman and Auli, Michael}, |
| journal={Advances in Neural Information Processing Systems}, |
| volume={33}, |
| year={2020} |
| } |
| ``` |
|
|
