Kyutai's Pocket TTS, ONNX int8: French and English bundles (mirror for Elabora)
This repository is an unmodified mirror kept by Elabora, an authoring tool for e-learning modules built around AI agents. Elabora downloads these files on the author's computer and checks their SHA-256 before using them. The upstream repository belongs to an individual (KevinAHM), who could rename or delete it at any time, so Elabora keeps a copy of exactly the bytes it was tested with.
Nothing was converted, retrained, recompressed or renamed: each file below is byte for byte the upstream file at the revision given, which the SHA-256 column lets anyone check. The credit for the model belongs to the people named under Upstream; Elabora claims none of it.
Upstream
- Mirrored repository:
KevinAHM/pocket-tts-onnx, revision58a6d00cf13d239b6748cb0769f35c580a8f606c, published by KevinAHM (an individual): an ONNX export of Kyutai's Pocket TTS, with int8 versions of the graphs, one folder per language bundle. Only the files Elabora uses of two bundles are mirrored, at the same paths as upstream:french_24landenglish_2026-04. The other bundles, the legacy root-level graphs, the scripts and the fp32 graphs of that repository are not copied. - Original model:
kyutai/pocket-tts, Pocket TTS by Kyutai (Manu Orsini, Simon Rouard, Gabriel De Marmiesse, Václav Volhejn, Neil Zeghidour, Alexandre Défossez), weights released under CC BY 4.0. The French bundle is Kyutai'sfrench_24l, the English oneenglish_2026-04. - Changes made upstream, by KevinAHM, not by Elabora: the weights exported to ONNX as separate graphs (text conditioner, flow language model in two parts, Mimi decoder and Mimi encoder), with int8 quantization of
flow_lm_main,flow_lm_flowandmimi_decoder(mimi_encoderandtext_conditionerstay fp32). - The bundles'
tokenizer.json(SentencePiece, by Kyutai) is not in this repository: Elabora downloads it fromkyutai/pocket-tts-without-voice-cloning, revision00eac05ed3d16bdc3f6b5d598874019c34a89214,languages/<bundle>/tokenizer.json. mimi_encoder.onnx(the voice encoder) is not needed to speak: Elabora downloads it only to build the voices of its own catalogue, once, on its developers' computers.
Files
Every file below is identical to the upstream file at the revision above. Upstream address of a file: https://huggingface.co/KevinAHM/pocket-tts-onnx/resolve/<revision>/<path>.
| File | Size (bytes) | SHA-256 |
|---|---|---|
onnx/english_2026-04/bos_before_voice.npy |
4224 | f46edf4f7007b7ba4ea58831f49d003e59e167b4641c44bb3addfe9231a780b1 |
onnx/english_2026-04/bundle.json |
24381 | bab643150f437f37df080a710520ff39ed9ebd9a339f8ebdc739f7eddfc28b3f |
onnx/english_2026-04/flow_lm_flow_int8.onnx |
9962530 | 3dd781ee5abee9e195320bf0106bebd6372a852b3b36352524ee78b40554635d |
onnx/english_2026-04/flow_lm_main_int8.onnx |
76341079 | f9bd8106b79a0192c1c43399ab938fb24900a95c1c599870d75a884e99000116 |
onnx/english_2026-04/mimi_decoder_int8.onnx |
22684077 | 3630450a3297a101792a6ac66619ebc70ab916b265e6220c2afaef8b1673f925 |
onnx/english_2026-04/mimi_encoder.onnx |
39768446 | 853e2ca623b8782d94c3745ec6133bfdff7ce33d9b11128bd29ea03f28d76e3d |
onnx/english_2026-04/text_conditioner.onnx |
16388344 | 4ecee995fb69f85c7a7493d11f7b5ee15d9950facc7ab3f5c9c49ef1e03847bb |
onnx/french_24l/bos_before_voice.npy |
4224 | 5759269802d691d49481b8e06d6a5649da85ea01135e270bd7641e3b091c7237 |
onnx/french_24l/bundle.json |
42235 | 8a5fe6c59985e3ccb5a6ccb1ffb2e84ac08488c5bfa704053618851436741427 |
onnx/french_24l/flow_lm_flow_int8.onnx |
9962530 | d340c549d5a1e0e7b88a0fb26fcae53c3b76486872f4418398a395bb1bc88701 |
onnx/french_24l/flow_lm_main_int8.onnx |
305144125 | 6130a6b98fae175147d82752263e250fb7b8483c1ef5373753ad335c16f4a129 |
onnx/french_24l/mimi_decoder_int8.onnx |
22684077 | b329ff3de3aa95455d2dee1cf371943dec269dd8711008017501ca129cb18d8c |
onnx/french_24l/mimi_encoder.onnx |
39768446 | 790a82fb8041e1035ec8097c6a277c834664a1f6619397d62f3bb46f71d2bfd4 |
onnx/french_24l/text_conditioner.onnx |
16388344 | 89d1b6ac55e618e42d2bd840438ce918bffb14960475ef091847ca39be641e65 |
To check a file: shasum -a 256 <path> (macOS, Linux) or certutil -hashfile <path> SHA256 (Windows).
Added by Elabora: LICENSE and this card.
Licence
Creative Commons Attribution 4.0 International (CC BY 4.0): the full legal code is in LICENSE (it is the upstream repository's onnx/LICENSE) and at https://creativecommons.org/licenses/by/4.0/legalcode. Attribution: Pocket TTS © Kyutai, licensed under CC BY 4.0 (original model), converted to ONNX by KevinAHM (ONNX files). Elabora changed nothing. The files are provided without warranty of any kind.
Prohibited use (from Kyutai's model card)
Use of our model must comply with all applicable laws and regulations and must not result in, involve, or facilitate any illegal, harmful, deceptive, fraudulent, or unauthorized activity. Prohibited uses include, without limitation, voice impersonation or cloning without explicit and lawful consent; misinformation, disinformation, or deception (including fake news, fraudulent calls, or presenting generated content as genuine recordings of real people or events); and the generation of unlawful, harmful, libelous, abusive, harassing, discriminatory, hateful, or privacy-invasive content. We disclaim all liability for any non-compliant use.
How Elabora uses it
Elabora's local voices speak narration lines on the author's computer with these graphs, run by onnxruntime-node, with a port of Kyutai's generation to TypeScript. Voices are built from recordings made available for speech synthesis; generated audio is marked as such.