phoonnx-supertonic

OpenVoiceOS mirror of the ONNX assets from Supertone/supertonic-3, used by the supertonic engine in phoonnx.

phoonnx mirrors all TTS model weights it uses under the OpenVoiceOS org on Hugging Face, so voices keep working even if an upstream repo is renamed, gated, or taken down. This is a straight mirror of the files referenced by phoonnx's supertonic voice index — no retraining, no modification to the weights.

Attribution

All weights and voice styles in this repository are produced by Supertone Inc. ("SuperTonic"). See the upstream repository for the original model card, demos, and the supertonic PyPI SDK: https://huggingface.co/Supertone/supertonic-3

Source code: https://github.com/supertone-inc/supertonic

License

The SuperTonic 3 weights are released by Supertone under an OpenRAIL-M license (license: openrail), same as upstream. Use of these weights is subject to the OpenRAIL-M use-based restrictions — see the upstream repository for the full license text.

Contents

• onnx/vector_estimator.onnx, onnx/text_encoder.onnx, onnx/duration_predictor.onnx, onnx/vocoder.onnx — the four ONNX graphs shared by all 310 voices (5 female + 5 male presets x 31 languages).
• onnx/tts.json, onnx/unicode_indexer.json — shared TTS config and grapheme indexer.
• voice_styles/{F1..F5,M1..M5}.json — the 10 per-speaker style embeddings. Language is selected at inference time; the style JSON only fixes the speaker identity.

Usage (via phoonnx)

from phoonnx.voice import TTSVoice

# any of the 310 "supertonic/<SPEAKER>/<lang>" voice ids from
# phoonnx/voice_index/supertonic.json, e.g. F1 speaking English
voice = TTSVoice.from_voice_id("supertonic/F1/en")
voice.say("A gentle breeze moved through the open window.")

phoonnx resolves the four shared ONNX graphs, the shared config/indexer, and the per-speaker style JSON straight from this repository.