DocsDownload
Menu

D57. Local models published by individuals are downloaded from Elabora's own copy on Hugging Face (19 Oct 2026)

Elabora’s optional local models are downloaded on first use from Hugging Face, each file pinned by revision and SHA-256 (the word aligners, D27; the transcriber, D46; the Pocket TTS voices, D25). A pin guarantees what the bytes are, not that they stay online: the French aligner, the English aligner, the transcriber and the voices were published by five individuals (XDcobra, jonatasgrosman, Xenova, istupakov, KevinAHM), any of whom can rename, empty or delete a repository, and Elabora would then lose a feature for every new user. On 19 October 2026 the founder agreed to copy them into Elabora’s own organization, elabora on Hugging Face, with the same bytes.

Decided:

  • The rule. A file published by an individual is downloaded from Elabora’s mirror of it; a file published by an organization (Kyutai’s tokenizers) stays with its authors, since an organization’s public repository is not one person’s to delete, unless a gate or a licence one day makes a mirror necessary. A file is mirrored only where its licence allows redistribution (Apache-2.0, MIT, CC BY 4.0…), with the attribution the licence asks; for a model that does not allow it, the options are our own export of an original that does, or another model. A new model published by an individual is mirrored before it is pinned.

  • The mirrors. One public repository per source repository, named after it, holding the pinned files at the same paths, byte for byte (the names, sizes and SHA-256 of the pins do not change), uploaded in one commit, which is the revision Elabora pins. Each has a model card: the upstream repository and revision, its authors, the licence and its full text (LICENSE), that it is an unmodified mirror, and the export script when there is one (the French aligner’s: XDcobra’s, kept as it is, and Elabora’s own export_wav2vec2.py, which gives the same word start for 98 % of the words, the others one frame apart, and is not the mirrored file).

    Mirror Upstream Licence Pinned revision
    elabora/wav2vec2-large-xlsr-53-french-onnx XDcobra/wav2vec2-large-xlsr-53-french-ONNX @ feb7081 (onnx/model_int8.onnx) Apache-2.0 fc9754d
    elabora/wav2vec2-large-xlsr-53-french jonatasgrosman/wav2vec2-large-xlsr-53-french @ 7c79e10 (vocab.json alone) Apache-2.0 dc59d3c
    elabora/wav2vec2-base-960h-onnx Xenova/wav2vec2-base-960h @ a19f851 (onnx/model_quantized.onnx, vocab.json) Apache-2.0 045b409
    elabora/parakeet-tdt-0.6b-v3-onnx istupakov/parakeet-tdt-0.6b-v3-onnx @ 8f23f0c (4 files) CC BY 4.0 (NVIDIA) bde3649
    elabora/pocket-tts-onnx KevinAHM/pocket-tts-onnx @ 58a6d00 (the two bundles’ files and voice encoders, onnx/…) CC BY 4.0 (Kyutai) 865fdd3

  • What did not change. The versions of the models (7c79e10-feb7081, a19f851, 8f23f0c-int8, 2026-04-58a6d00), their names, sizes and SHA-256: installed files are recognized by those (D55), not by the address they came from, so every model already installed stays current and none is downloaded again; the pin moved, the model did not (test/unit/model-mirrors.test.ts). ELABORA_ALIGNER_MODELS_URL, ELABORA_TRANSCRIBER_MODELS_URL and ELABORA_VOICE_MODELS_URL still replace https://huggingface.co. The upstream of each constant is named in a comment beside it. Because the mirrors are on Hugging Face, the hosts the app contacts and the privacy page are unchanged.

  • A guard. The same test fails when a pinned file is not from elabora/ or Kyutai, and when a mirror could change the bytes it serves (the SHA-256 is still checked).

  • Not mirrored. Kyutai’s two tokenizer.json (kyutai/pocket-tts-without-voice-cloning @ 00eac05, 490 KB): an organization, CC BY 4.0, open to everyone today, though its model card carries the prohibited-use form that its sibling kyutai/pocket-tts enforces as a gate. The recordings of kyutai/tts-voices that scripts/pocket-voices.ts reads to make the catalogue (CC0 and CC BY, each pinned by URL and SHA-256 in sources.json) are used by developers, once, never by an author’s computer. Nothing else is downloaded from Hugging Face.

  • One licence to watch. Xenova’s repository, the English aligner’s, declares no licence of its own: its files are a conversion and a quantization of facebook/wav2vec2-base-960h (Apache-2.0), mirrored under the original’s licence and said so on the card. If its publisher ever objected, the answer is our own export of the original, which changes the file and so needs the English start and end shifts measured again (D27).

Revisit if: Kyutai gates its no-cloning repository or changes its licence (mirror the tokenizers, three lines), a mirror’s organization is renamed (the pins name it), a mirrored model’s licence is withdrawn or turns out not to allow redistribution, or Hugging Face’s terms stop serving public models without an account.