#Supertonic: Lightning-Fast, On-Device Text-to-Speech

A new open-source text-to-speech engine can turn text into speech on your own machine, faster than you can listen to it, without ever calling a cloud API.

#What it is

Supertonic is a local text-to-speech system built by Supertone Inc. and powered by ONNX Runtime. You feed it text, it produces audio, and the whole thing runs on your own device. There is no cloud round trip, no API call to a third-party server, and no upload of your content. The design goal is simple: make on-device TTS fast enough and light enough to be practical.

The current generation, Supertonic 3, is roughly 99 million parameters across its public ONNX assets, according to the GitHub README. That is small compared to many open TTS models in the 0.7B to 2B range, which matters for download size, startup time, and memory footprint. Supertonic 3 also expanded from 5 to 31 languages, including Vietnamese, and ships examples for Python, Node.js, Java, C++, C#, Go, Swift, iOS, Rust, and Flutter.

#Why it matters

• Local TTS can cut API costs if you generate a lot of voiceover, because you are not paying per character or per request.
• Your text stays on your device, which is a meaningful privacy improvement for sensitive or unreleased content.
• High throughput means you can read web pages, batch-generate audio, or power an assistant without waiting on a network.
• The breadth of platform examples means you can embed it into a browser, a mobile app, or an edge device without rewriting your stack.
• Vietnamese appears in the official benchmark table, which makes it worth testing for anyone producing Vietnamese-language content.

#How it works

Under the hood, Supertonic runs inference through ONNX Runtime. On a modern machine, the speed claims are substantial. According to the Hugging Face performance table, an M4 Pro CPU reaches 912 to 1,263 characters per second with 2-step inference, and the same machine on WebGPU reaches 996 to 2,509 characters per second. An RTX 4090 is listed at 2,615 to 12,164 characters per second. These are project-reported numbers, not independent benchmarks, but the pattern is clear: the model generates audio far faster than realtime.

Beyond raw speed, the repo is not just a model file. It includes a Python SDK, a local HTTP server, an OpenAI-compatible /v1/audio/speech endpoint, a browser WebGPU path, and examples for a dozen runtimes. Text normalization handles currency, phone numbers, dates, and technical units, which is where many TTS systems stumble.

#Caveats

The open repository focuses on fixed voices, not zero-shot voice cloning. If you want custom voices, Supertone offers a Voice Builder and managed products separately. The speed numbers are project-reported and should be tested independently on your target hardware before production use. The model card on Hugging Face still lists an older 66M parameter figure, while the current README says 99M, so documentation can drift between versions. And while Vietnamese is in the benchmark table, accent quality and naturalness still need a real listening test on your specific content.

#Who it's for

Supertonic is for developers, content creators, and teams that generate enough voiceover to care about cost, latency, or privacy, and who are willing to install and benchmark a local model rather than rely on a managed API.

If you need local, private, fast TTS that you can embed across platforms, Supertonic is one of the most practical open options available right now, but verify the numbers on your own hardware before committing.

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