**Supertonic** is a lightning-fast, on-device text-to-speech system designed for local inference with minimal overhead. Powered by ONNX Runtime, it runs entirely on your device—no cloud, no API calls, no privacy concerns.
- **2026.01.22** - **[Voice Builder](https://supertonic.supertone.ai/voice_builder)** is now live! Turn your voice into a deployable, edge-native TTS with permanent ownership.
- **2026.01.06** - 🎉 **Supertonic 2** released with 5-language support. The v2 code path is preserved on the [`release/supertonic-2`](https://github.com/supertone-inc/supertonic/tree/release/supertonic-2) branch.
- **2025.12.10** - Added `supertonic` PyPI package! Install via `pip install supertonic`. For details, visit [supertonic-py documentation](https://supertone-inc.github.io/supertonic-py)
- **2025.12.10** - Added [6 new voice styles](https://huggingface.co/Supertone/supertonic/tree/b10dbaf18b316159be75b34d24f740008fddd381) (M3, M4, M5, F3, F4, F5). See [Voices](https://supertone-inc.github.io/supertonic-py/voices/) for details
- **2025.12.08** - Optimized ONNX models via [OnnxSlim](https://github.com/inisis/OnnxSlim) now available on [Hugging Face Models](https://huggingface.co/Supertone/supertonic)
- **2025.11.24** - Added Flutter SDK support with macOS compatibility
Supertonic 3 is designed for practical on-device inference: compact enough to run locally, while staying competitive with much larger open TTS systems.
Across measured languages, Supertonic 3 stays within a competitive WER/CER range against much larger open TTS models such as VoxCPM2, while preserving a lightweight on-device deployment path. Asterisked languages use CER; the others use WER.
Compared with Supertonic 2, Supertonic 3 reduces repeat and skip failures, improves speaker similarity across the shared-language set, and expands language coverage from 5 to 31 languages. It keeps the v2-compatible public ONNX interface, so existing integrations can move to v3 with the same inference contract.
<img src="img/metrics/runtime_cpu_gpu_latency_memory.png" alt="Supertonic CPU runtime compared with GPU baselines">
</p>
Supertonic 3 runs fast on CPU, even compared with larger baselines measured on A100 GPU, and uses substantially less memory. The open-weight fixed-voice setting does not require a GPU, which makes local, browser, and edge deployment much easier.
At about 99M parameters across the public ONNX assets, Supertonic 3 is much smaller than 0.7B to 2B class open TTS systems. The smaller model size is a practical advantage for download size, startup time, and on-device inference.
## Demo
> **Try it now**: Experience Supertonic in your browser with our [**Interactive Demo**](https://huggingface.co/spaces/Supertone/supertonic-3), or get started with pre-trained models from [**Hugging Face Hub**](https://huggingface.co/Supertone/supertonic-3)
### Raspberry Pi
Watch Supertonic running on a **Raspberry Pi**, demonstrating on-device, real-time text-to-speech synthesis:
Turns any webpage into audio in under one second, delivering lightning-fast, on-device text-to-speech with zero network dependency—free, private, and effortless:
> 🎧 **View audio samples more easily**: Check out our [**Interactive Demo**](https://huggingface.co/spaces/Supertone/supertonic-3) for a better viewing experience of all audio examples
> **Note:** These samples demonstrate how each system handles text normalization and pronunciation of complex expressions **without requiring pre-processing or phonetic annotations**.
The following papers describe the core technologies used in Supertonic. If you use this system in your research or find these techniques useful, please consider citing the relevant papers:
### SupertonicTTS: Main Architecture
This paper introduces the overall architecture of SupertonicTTS, including the speech autoencoder, flow-matching based text-to-latent module, and efficient design choices.
```bibtex
@article{kim2025supertonic,
title={SupertonicTTS: Towards Highly Efficient and Streamlined Text-to-Speech System},
author={Kim, Hyeongju and Yang, Jinhyeok and Yu, Yechan and Ji, Seunghun and Morton, Jacob and Bous, Frederik and Byun, Joon and Lee, Juheon},
journal={arXiv preprint arXiv:2503.23108},
year={2025},
url={https://arxiv.org/abs/2503.23108}
}
```
### Length-Aware RoPE: Text-Speech Alignment
This paper presents Length-Aware Rotary Position Embedding (LARoPE), which improves text-speech alignment in cross-attention mechanisms.
```bibtex
@article{kim2025larope,
title={Length-Aware Rotary Position Embedding for Text-Speech Alignment},
author={Kim, Hyeongju and Lee, Juheon and Yang, Jinhyeok and Morton, Jacob},
journal={arXiv preprint arXiv:2509.11084},
year={2025},
url={https://arxiv.org/abs/2509.11084}
}
```
### Self-Purifying Flow Matching: Training with Noisy Labels
This paper describes the self-purification technique for training flow matching models robustly with noisy or unreliable labels.
```bibtex
@article{kim2025spfm,
title={Training Flow Matching Models with Reliable Labels via Self-Purification},
author={Kim, Hyeongju and Yu, Yechan and Yi, June Young and Lee, Juheon},
This project's sample code is released under the MIT License. - see the [LICENSE](https://github.com/supertone-inc/supertonic?tab=MIT-1-ov-file) for details.
The accompanying model is released under the OpenRAIL-M License. - see the [LICENSE](https://huggingface.co/Supertone/supertonic-3/blob/main/LICENSE) file for details.
This model was trained using PyTorch, which is licensed under the BSD 3-Clause License but is not redistributed with this project. - see the [LICENSE](https://docs.pytorch.org/FBGEMM/general/License.html) for details.
