Supertonic 3 (int8) for Any Text

The on-device speech files of the Any Text apps (iOS and Android): reading a text aloud and saying a single word. The apps download these files once and run them on the phone with ONNX Runtime. No text or audio leaves the phone.

This is not an original model. All weights are Supertone Inc.'s Supertonic 3, released under the BigScience Open RAIL-M licence (copy in LICENSE).

Sources (pinned)

Files	Source	Revision
`onnx/tts.json`, `onnx/unicode_indexer.json`, `voice_styles/*.json`	Supertone/supertonic-3	`3cadd1ee6394adea1bd021217a0e650ede09a323`
`onnx/*.int8.onnx`, `sherpa/unicode_indexer.bin`, `sherpa/voice.bin`	csukuangfj2/sherpa-onnx-supertonic-3-tts-int8-2026-05-11 (the k2-fsa sherpa-onnx project)	`cca5a0e6c96e1d2c720986bf7e75fcc81dee3ae4`

Modifications (Open RAIL-M §4c)

onnx/duration_predictor.int8.onnx, onnx/text_encoder.int8.onnx, onnx/vector_estimator.int8.onnx and onnx/vocoder.int8.onnx are 8-bit quantized versions of Supertone's onnx/*.onnx, made by the sherpa-onnx project. The weights were changed by quantization only.
sherpa/unicode_indexer.bin and sherpa/voice.bin hold the same data as onnx/unicode_indexer.json and the ten voice_styles/*.json, converted to the binary format sherpa-onnx reads (voice.bin: speaker ids 0–4 = F1–F5, 5–9 = M1–M5).
All other files are unchanged copies.

manifest.json lists every file with its size and SHA-256. The apps check each file against it after downloading.

Use restrictions

Use of these files is subject to paragraph 5 and Attachment A of the Open RAIL-M licence. The Any Text terms of use pass these restrictions on to the app's users, and the app says that its voices are machine-generated.

Credits

Downloads last month: -; Downloads are not tracked for this model. How to track