Supertone releases Supertonic 3, a 99M on-device TTS in 31 languages
Supertonic 3 is a roughly 99M ONNX TTS for local and browser use across 31 languages; OpenRAIL-M model, MIT sample code.
Supertone released Supertonic 3, a text-to-speech model of about 99M parameters shipped as ONNX assets. The card says it runs locally with ONNX Runtime with no cloud call and no GPU requirement, which suits browser, edge and local use. It supports 31 languages, up from 5 in Supertonic 2, including English, Japanese, Korean, German, French, Hindi and Vietnamese. Expression tags such as laughs, breaths and sighs are supported.
The card says CPU speed beats larger baselines measured on an A100 and that memory use is much lower, and describes reading accuracy as "a competitive WER/CER range" against larger open models such as VoxCPM2. The exact figures appear only in charts, not in the text, so none are repeated here.
The open package includes only fixed preset voice styles. Voice cloning is not part of it; custom voice styles come from a separate tool, Supertonic Voice Builder, and purchased styles are tied to that tool. Licensing is mixed: sample code is MIT and the model is OpenRAIL-M.
The card gives no release date. The date used is the Hugging Face repo creation date.
Small, local and honest about what is not included.
Sources
Editor's first pass (AI-assisted)
- Real 4: ONNX assets are public and run locally without a GPU, per the card.
- Useful 3.5: Fast local synthesis is useful, but the open package has fixed preset voices only.
- Notable 3.5: Over 1,000 likes on Hugging Face for a very small model.
- Novel 3: Small-model efficiency is the point; the approach is incremental on Supertonic 2.
Comments