Supertonic

Supertonic goes where big TTS models cannot: about 99M parameters running through ONNX Runtime entirely on-device, synthesizing 44.1kHz audio faster than real time on plain CPUs, phones, browsers via WebGPU, and even Raspberry Pi boards and e-readers. Version 3, released May 2026, expanded coverage from 5 to 31 languages and ships ten preset voice styles plus inline expression tags for things like laughter and breaths. Integration is the selling point: ready-made SDKs cover Python, Node.js, browser JavaScript, Java, C++, C#, Go, Rust, Swift, and Flutter, a pip package installs the engine in one line, and a local HTTP server exposes OpenAI-compatible endpoints. Sample code is MIT while the model weights carry an OpenRAIL-M license, with everything downloadable from Hugging Face. One caveat: developer Supertone ended official support in July 2026 and its hosted Voice Builder is being retired, so the repo is frozen in a working state rather than evolving. For private, offline speech on weak hardware it remains hard to beat.