An on-device text-to-speech (TTS) model released by Supertone, a Korean voice AI company. It is designed to complete all speech synthesis on the local device, requiring neither an internet connection nor an API key. Its defining trait is its small size: just 99M (99 million) parameters. Despite being orders of magnitude lighter than TTS models in the 0.7B–2B range, it reads aloud in 31 languages, including Japanese, from a single model. It uses ONNX Runtime for inference and runs on CPU alone, with no GPU required. Sample implementations are provided for 11 runtimes — Python, Node.js, browser (WebGPU), Java, C++, C#, Go, Swift, iOS, Rust, and Flutter — so it fits almost any integration target.
Note that the official repository carries a notice dated July 23, 2026, announcing that it will be archived and that official support has ended. Voice Builder, a related service, will also become inaccessible on August 31, 2026. Anyone considering adoption should factor this in.
Key Features
- Fully on-device operation: Because synthesis runs entirely locally, your text is never sent to an external server. It works offline, and there are no API keys to manage and no usage-based billing. This suits privacy-sensitive use cases well
- Lightweight 99M-parameter design: While typical large TTS models range from hundreds of millions to billions of parameters, Supertonic fits in 99M. Supertone reports an average real-time factor (RTF) of 0.3x even on a Raspberry Pi — roughly one-third of the audio’s duration to generate it
- 31 languages in a single model: Supports 31 languages including English, Japanese, Korean, German, French, Spanish, Arabic, Hindi, and Vietnamese. There is no need to swap models per language, and when the language cannot be determined, a language-agnostic mode (
lang="na") is available - Sample implementations for 11 runtimes: Official sample code exists for Python, Node.js, browser (WebGPU), Java, C++, C#, Go, Swift, iOS, Rust, and Flutter. It can be embedded in mobile apps, desktop apps, or web front ends alike
- Easy install via
pip: In a Python environment, a singlepip install supertonicis enough. To simply try it out, a demo Space on Hugging Face runs in the browser - Published accuracy benchmarks: WER / CER comparison results on Minimax-MLS-test are published, showing reading accuracy on par with far larger models
Pricing
| Plan | Price | What you get |
|---|---|---|
| Model itself (self-hosted) | Free | Download the model weights and sample code and run them on your own device or server. No usage-based charges |
| Supertone Play | Check the official site (Starter includes a free first month) | Supertone’s web and desktop product. 263 preset voices with emotional expression, plus API integration |
Pricing reflects information as of August 2026. Please check the official site for the exact figures of Supertone Play’s Creator / Pro plans. For the latest information, see the official site.
Licensing is split in two. The sample code and runtime bindings are MIT licensed, while the model weights are under OpenRAIL-M. OpenRAIL-M includes use-based restrictions such as “no harm” and “no impersonation without consent,” so read the license text carefully before any commercial use.
Pros & Cons
✅ Pros
- Since nothing is transmitted to a server, it can be used for reading confidential drafts or text containing personal information
- Zero running cost. Unlike usage-based cloud TTS, reading large volumes does not increase your bill
- The model is light enough to run at practical speed on Raspberry Pi-class hardware, making it easy to put on embedded devices and kiosks
- One model covers 31 languages, which simplifies model management in multilingual apps
- Samples for 11 runtimes keep the implementation cost of embedding it into an existing app low
⚠️ Cons
- The official repository is announced as pending archive with support ended, so future features and bug fixes cannot be expected
- Voice Builder (zero-shot voice cloning) becomes inaccessible on August 31, 2026
- The model weights are OpenRAIL-M, not fully unrestricted. You need to verify the use-based restrictions
- Compared with the top tier of commercial cloud TTS, the range of control over emotion and fine intonation is limited
- Adoption requires at least basic programming knowledge. This is not a product that works through a GUI alone
Comparison with Similar Services
| Criteria | Supertonic | Kokoro TTS | Piper | ElevenLabs |
|---|---|---|---|---|
| Where it runs | On-device | On-device | On-device | Cloud (API) |
| Model size | ~99M parameters | ~82M parameters | Lightweight (per-voice models) | Undisclosed |
| Languages | 31 (single model) | English-centric plus others | Multilingual (per-language models) | Multilingual |
| Japanese | Supported | Supported | Supported | Supported |
| Price | Free (self-hosted) | Free | Free | Usage-based / subscription |
| License | Code MIT / model OpenRAIL-M | Apache-2.0 | MIT | Commercial service |
| Development status | Pending archive | Active | Active | Active |
Who Is It For
- Developers building internal, medical, or legal applications where the text to be read aloud cannot leave the organization
- Teams that need to synthesize large volumes of text and find usage-based cloud TTS pricing uneconomical
- Anyone implementing speech output on modest hardware such as Raspberry Pi or embedded devices
- Developers of multilingual apps who want to avoid switching TTS engines per language
- People who want to study a lightweight local TTS implementation, or build a prototype that must work offline
Summary
Supertonic is an implementation-focused TTS model that delivers on-device reading in 31 languages at just 99M parameters. Requiring neither an internet connection nor an API key pays off in work with strict privacy requirements and in high-volume reading where cost matters. From the perspective of actually embedding it into an app, the samples for 11 runtimes are a significant advantage.
That said, the official repository has announced a pending archive and the end of support. Avoid adopting it on the assumption of future updates; it makes sense when you can accept using what is finished today, as is. If you need ongoing support or zero-shot voice cloning, consider commercial products such as Supertone Play, or other on-device TTS projects that are still under active development.