Supertonic-3
This first-party Soniqo page documents Supertonic-3 from the local speech-swift / speech-core implementation. Hugging Face bundles are linked below after the integration notes.
Internal Page First
Landing cards and docs menus now point here first; source model and bundle links remain available from this page.
At a Glance
| Model | Supertonic-3 |
|---|---|
| Role | G2P-free multilingual text-to-speech |
| Backend | CoreML (Apple Neural Engine / GPU) and LiteRT |
| Output | 44.1 kHz mono Float32 PCM |
| Languages | 31 languages plus a neutral na tag |
| License | OpenRAIL-M weights, MIT code |
| Status | Ready for CoreML; LiteRT reference implementation in speech-core |
| Source | Supertone Supertonic-3 |
| Swift product | SupertonicTTS |
| CLI / runtime | Shared TTS pipeline / server integration; LiteRT C++ wrapper for edge runtimes |
Use
The snippet below mirrors the current speech-swift API or command exposed by the repo.
import SupertonicTTS
let tts = try await SupertonicTTSModel.fromPretrained()
let pcm = try tts.synthesize(
text: "Hello from an on-device voice.",
voiceId: "F1",
language: "en"
)Model Links
Implementation Notes
- No espeak, phonemizer, or lexicon: text is NFKD-normalized and mapped through a Unicode index table.
- Download uses explicit CoreML package globs for the four graphs plus tokenizer, config, and voice style JSON files.
- The Apple export uses dynamic latent length; the current LiteRT path uses fixed graph shapes and chunks longer text.
- Voices are precomputed style presets F1-F5 and M1-M5; on-device cloning is out of scope because the style extractor is not released.