// the find
supertone-inc/supertonic
Lightning-Fast, On-Device, Multilingual TTS — running natively via ONNX.
Supertonic is a 99M-parameter on-device TTS system from Supertone Inc., running via ONNX Runtime across 31 languages with no cloud dependency. It targets edge deployments — Raspberry Pi, e-readers, browsers via WebGPU — where latency and privacy matter more than voice variety. If you need TTS that works offline and ships in 11+ languages without a GPU, this is one of the few options actually worth evaluating.
SDK breadth is real: Python, Node, C#, Go, Rust, Swift, iOS, Flutter, C++ — each with a working example and not just a stub. The ONNX backend means you're not locked to one runtime or platform. Text normalization for financial expressions, phone numbers, and technical units is handled natively, which is something larger cloud APIs consistently get wrong. The `lang="na"` fallback for unknown input language is a practical touch that avoids forcing callers to detect language themselves.
Voice variety is thin — a handful of preset voices, no zero-shot cloning in the open-weight repo (that lives behind Voice Builder, a separate web service). The WER/CER numbers show some languages are noticeably weaker: Vietnamese at 4.49, Finnish at 5.40, Japanese at 4.61, all worse than the competition they benchmarked against. The model license is OpenRAIL-M, not MIT — commercial use has acceptable-use restrictions the README buries in a footnote. Git LFS is required to pull the model weights from Hugging Face, which is a friction point in CI or containerized builds.