threebobss/supertonic-3-tts-int8-sherpa

Supertonic-3 (int8, sherpa-onnx layout)

A re-host of the sherpa-onnx project's int8 export of Supertonic-3 by Supertone Inc., a multilingual on-device text-to-speech model (31 languages, ten preset voices). The files are pinned here so that an app can download them from a fixed commit.

What these files are

• Model: Supertonic-3 by Supertone Inc., released under the OpenRAIL-M licence (see LICENSE, copied verbatim from Supertone/supertonic-3 at commit 3cadd1ee6394adea1bd021217a0e650ede09a323).
• Modification notice: these files are not Supertone's original fp32 model. They are the sherpa-onnx project's int8 quantization of Supertonic-3, published as sherpa-onnx-supertonic-3-tts-int8-2026-05-11.tar.bz2 in the tts-models release of k2-fsa/sherpa-onnx. voice.bin and unicode_indexer.bin are sherpa-onnx's binary packing of Supertone's voice_styles/*.json and unicode indexer. They are re-hosted here unchanged: every file is byte-identical to that release (sha256 below).
• Runtime: sherpa-onnx (Apache-2.0), OfflineTtsSupertonicModelConfig. The voice is picked with a speaker id: voice.bin holds the ten styles in sorted order, so sid 0–4 = F1–F5 and sid 5–9 = M1–M5. The language goes in GenerationConfig.extra["lang"] (for example de).

Use restrictions

The OpenRAIL-M licence's Attachment A use restrictions apply to every use of these files and of anything derived from them. They are in full at the end of LICENSE. Anyone who uses or redistributes the model must pass these restrictions on to their own users.

Machine-generated content: audio produced with this model is computer-generated speech. Under Attachment A (e), it must not be presented as human speech, and content made with it must be clearly marked as machine-generated where it is shared.

Files

File	Bytes	sha256
duration_predictor.int8.onnx	3,700,147	c3eb91414d5ff8a7a239b7fe9e34e7e2bf8a8140d8375ffb14718b1c639325db
text_encoder.int8.onnx	36,416,150	c7befd5ea8c3119769e8a6c1486c4edc6a3bc8365c67621c881bbb774b9902ff
vector_estimator.int8.onnx	78,400,833	20cd86fa5c6effedfda0e7cffe5b0569ca401c440a0c3a1d72bf39286c0db3fd
vocoder.int8.onnx	25,991,073	e923d60f53f95eb1ce235f1dc33ec56d9c057823c96fa6f8acf98f32b0da6152
voice.bin	517,168	67d5209b0ee8ce6c74105ffbe12fe6a7628aea3b4ba2fcb308a4a67938a93ce8
unicode_indexer.bin	262,144	8402ca48e5189a8950138580b0fff64db6f072f24ac07cd54ba8b2fbb9883b30
tts.json	8,253	42078d3aef1cd43ab43021f3c54f47d2d75ceb4e75f627f118890128b06a0d09

Total: 145,295,768 bytes.

Credits

• Supertonic-3: Supertone Inc. (model, code).
• int8 export and runtime: the sherpa-onnx project (k2-fsa).