KorvaTTS — Vietnamese-first, on-device text-to-speech

🇻🇳 Tiếng Việt

KorvaTTS is an open-source text-to-speech system focused on Vietnamese with natural code-switching (Vietnamese sentences that contain English words, brand names, tech terms). It runs entirely on-device through ONNX Runtime: no GPU required, no API calls, 44.1 kHz output.

The model is our own re-implementation of the 99M-parameter Supertonic 3 architecture (flow-matching TTS), trained from scratch on Vietnamese data, with a vocoder based on BlueCodec. Upstream Supertonic is being archived by its authors; KorvaTTS continues the line with a Vietnamese focus and an open roadmap toward voice cloning and training.

Highlights

• 🇻🇳 Vietnamese-first — trained on Vietnamese audiobook speech plus a private code-switching set
• 🔀 Code-switching — English words inside Vietnamese sentences are read naturally
• ⚡ Fast on CPU — real-time factor well below 1 on a laptop CPU; GPU optional
• 🔊 44.1 kHz WAV output, no external upsampler
• 🧩 Drop-in compatible with Supertonic 3 runtimes (see Compatibility)
• 🪪 Apache-2.0 for both code and weights

Roadmap

Phase 2 starts once the phase-1 release has a healthy community around it — our goal is 1,000 GitHub stars ⭐ . Star the repo to move the roadmap forward, and follow the issues tab / discussions for progress.

Language support

• Primary: Vietnamese (lang="vi"), including English words embedded in Vietnamese text.
• English (lang="en"): part of the training data is English, so plain English works, but we do not guarantee English quality matches — or exceeds — the original Supertonic 3 English voices. Vietnamese is what this model is built for.
• Other languages: the Supertonic 3 tag set (31 languages) is still accepted by the text encoder, but those languages were not in the training data; quality is untested.

Training data

• PhoAudiobook — 941 hours of curated Vietnamese audiobook speech, introduced in Zero-Shot Text-to-Speech for Vietnamese (ACL 2025).
• A private, author-collected Vietnamese/English code-switching set used to improve mixed-language reading. This set is not released.

Installation

pip install korvatts

Requires Python ≥ 3.10. For GPU inference install onnxruntime-gpu (CUDA) instead of onnxruntime.

Quick start

from korvatts import TTS

tts = TTS()  # first run downloads weights + voices from Hugging Face
print(tts.list_voices())

wav, duration = tts.synthesize(
    "Hôm nay team mình sẽ demo tính năng text-to-speech mới.",
    voice="khanh_vy",
    lang="vi",
    total_steps=32,  # default; best quality. Lower (8-16) for faster synthesis
    speed=1.0,       # 0.5 .. 2.0
)
tts.save_audio(wav, "output.wav")

Command line:

korvatts voices
korvatts synth "Xin chào, đây là KorvaTTS." -v khanh_vy -o hello.wav

Use local assets instead of downloading: pass TTS(assets_dir="path/to/assets") or set KORVATTS_ASSETS_DIR. The directory must contain onnx/ and voice_styles/.

Voices

10 bundled voices — 5 female, 5 male. Listen to every voice on the Hugging Face model card.

Compatibility with Supertonic

The phase-1 weights keep the exact ONNX graph signatures and voice-style JSON format of Supertonic 3. You can copy the KorvaTTS onnx/ and voice_styles/ folders into any Supertonic 3 runtime (Python, Node.js, browser/WebGPU, Rust, C++, Swift, Flutter, …) and they will work unchanged.

This compatibility is intentional for phases 1–2. Phase 3 will change the architecture and will not be loadable by Supertonic runtimes. Version tags on Hugging Face will make it clear which checkpoints are compatible.

Licenses

Portions of the inference code are adapted from Supertonic (MIT, © Supertone Inc.); the vocoder follows BlueCodec from BlueTTS (MIT). See NOTICE. KorvaTTS is not affiliated with Supertone Inc.

Acknowledgements & citation

KorvaTTS builds on the work of others. If you use it, please also cite them:

• Supertonic / SupertonicTTS — architecture and reference implementation. github.com/supertone-inc/supertonic
• BlueTTS / BlueCodec — vocoder design. github.com/maxmelichov/BlueTTS
• PhoAudiobook — Vietnamese training corpus. huggingface.co/datasets/thivux/phoaudiobook

@misc{korvatts2026,
  title        = {KorvaTTS: Vietnamese-first on-device text-to-speech},
  author       = {dogenthq},
  year         = {2026},
  howpublished = {\url{https://github.com/dogenthq/KorvaTTS}}
}

@article{kim2025supertonictts,
  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}
}

@misc{melichov2025bluetts,
  title        = {BlueTTS},
  author       = {Melichov, Max},
  year         = {2025},
  howpublished = {\url{https://github.com/maxmelichov/BlueTTS}}
}

@inproceedings{vu2025phoaudiobook,
  title     = {Zero-Shot Text-to-Speech for Vietnamese},
  author    = {Vu, Thi and Nguyen, Linh The and Nguyen, Dat Quoc},
  booktitle = {Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Short Papers)},
  year      = {2025}
}

Support the project

KorvaTTS is free and open source. Training runs and voice recordings cost real money — if this project is useful to you, a small donation goes directly into the voice cloning and training work in phase 2. Scan the VietQR code below with any Vietnamese banking app (international options coming later).

Contributing

Issues and pull requests are welcome. Run the checks locally:

pip install -e ".[dev]"
ruff check .
pytest        # model tests auto-skip unless ./assets exists