supertonic

Most production text to speech still depends on a cloud round trip: you send text to an API, pay per character, and wait on the network before audio comes back. That model breaks down for offline apps, privacy sensitive contexts, and anything that needs to start speaking instantly. supertonic, from Supertone, is a text to speech engine built to run entirely on the device that needs the speech. It ships a compact model of roughly 99 million parameters, far smaller than the 0.7 billion to 2 billion parameter systems it positions itself against, and it covers 31 languages including English, German, Spanish, French, Japanese, Korean, and Arabic. The point is simple: usable narration that runs locally on a CPU, in a browser, or on a phone, with no key and no network.

supertonic runs through ONNX Runtime, which is what lets the same model execute across CPU, WebGPU in the browser, and mobile without a bespoke runtime per platform. Architecturally the system combines a speech autoencoder with a flow matching module that maps text to latent representations, an approach chosen to keep the model small while staying fast. The project reports an average real time factor around 0.3 on modest hardware such as a Raspberry Pi and e readers, meaning audio generates faster than it plays, and it claims it can turn an entire web page into audio in under a second. Output is 44.1kHz 16 bit WAV. The breadth of provided SDKs is the other distinctive trait: ready implementations exist for Python, Node, the browser, Java, C++, C#, Go, Swift, iOS, Rust, and Flutter.

The clearest fit is any product where latency, cost per call, or privacy rules out a cloud TTS service. Browser extensions that read articles aloud, e book and e reader apps, voice chatbots that need an instant local response, and accessibility tools all benefit from synthesis that starts immediately and never leaves the device. The project already lists real integrations along these lines, including read aloud extensions, an e book reader, and avatar generators. For developers, the wide SDK coverage means you can prototype in Python and ship the same model inside a Swift iOS app, a Rust binary, or a Flutter mobile build without switching engines. Edge and embedded scenarios, where there may be no reliable connection at all, are where the on device design pays off most clearly.

The honest caveats matter here. The open variant ships a fixed voice and has no built in voice cloning: custom voices require Supertone's separate Voice Builder tool, which produces profile files rather than being part of the open repository. Licensing is split, and this is easy to miss. The sample code is MIT, but the model weights are released under OpenRAIL-M, an open weight license that carries use restrictions, so this is not fully OSI open and commercial users need to read the clauses. Quality is uneven across the 31 languages: the project itself reports a Japanese character error rate of 4.61 percent, so some languages clearly perform better than others. The repository also had 78 open issues at the time of writing, a normal sign of an active young project but worth weighing for production planning.

Against the commercial path, Supertone Play and the hosted API, supertonic trades managed cloud convenience and built in cloning for control, zero per call cost, and full on device privacy. Adopt it when you are building edge, mobile, or browser TTS and latency or data residency genuinely matter, when you can live with the provided voice or are willing to use Voice Builder separately, and when the OpenRAIL-M terms fit your use. It is too early to lean on if you need out of the box voice cloning, if you want broadcast grade consistency in a less supported language, or if your legal team cannot accept open weight use restrictions for a commercial product. For local first, latency sensitive speech, though, it is a strong and unusually portable option worth testing directly.