supertonic-3

Part of the Supertonic 3 MLX collection.

Supertonic 3 (MLX)

Apple MLX graph-runtime conversion of Supertone/supertonic-3, a compact multilingual text-to-speech model distributed upstream as ONNX assets.

The model can be run programmatically using the open-source MLX runtime or through OpenVox, a desktop application for generating speech, reading selected text aloud, creating audiobooks, and exposing local TTS models to other applications and AI agents.

TL;DR

Use with OpenVox

OpenVox provides a graphical desktop interface for running Supertonic 3 locally without requiring a Python environment or command-line setup.

With OpenVox, the model can be used for:

• Generating speech from text
• Reading selected text aloud from other applications
• Creating and exporting long-form audio and audiobooks
• Choosing from a large multilingual voice library
• Adjusting voice, language, speed, and generation parameters
• Processing text locally after the required model files are downloaded
• Accessing local TTS capabilities through an API or MCP-compatible AI agents

OpenVox is especially useful for users who want an integrated, on-device TTS workflow while retaining the ability to use the underlying MLX model directly.

OpenVox is a third-party application and is not affiliated with or endorsed by Supertone.

Quick Start

OpenVox

Install OpenVox, download Supertonic 3 from its model library, select a voice and language, and enter or import the text you want to convert to speech.

Once downloaded, the model can be used locally without sending the input text to a cloud inference service.

Python and MLX

hf download mlx-community/supertonic-3 --local-dir ./models/supertonic-3

git clone https://github.com/ailuntx/supertonic-mlx.git
cd supertonic-mlx

python -m venv .venv
.venv/bin/pip install mlx soundfile numpy

.venv/bin/python scripts/infer_mlx.py \
  --model ./models/supertonic-3 \
  --text "Supertonic 3 is running with MLX." \
  --lang en \
  --voice M1 \
  --total-step 8 \
  --output output.wav

Model Layout

supertonic-3/
├── README.md
├── mlx_manifest.json
├── graphs/
├── weights/
└── voice_styles/

Conversion Notes

Validation

The MLX graph runtime has been checked against ONNX Runtime using the official assets. Per-stage maximum absolute errors are approximately 1e-5.

The Hugging Face Space API has also successfully returned generated audio with real wall-time status reporting.

Intended Use

This model is intended for text-to-speech generation, accessibility, read-aloud applications, audiobook creation, prototyping, local AI workflows, and research.

Users are responsible for ensuring that generated audio is used lawfully and ethically. Do not use the model to impersonate individuals, deceive listeners, or create misleading audio without appropriate disclosure and consent.

License

The model license follows the upstream Supertonic 3 model card (openrail).

OpenVox and the supertonic-mlx runtime may be distributed under their own respective software licenses. Consult each project’s documentation before redistribution or commercial integration.

Attribution

This repository contains an MLX-compatible conversion of the upstream Supertonic 3 assets. The original model architecture, research, and model assets were created by the Supertonic authors.

OpenVox provides an independent desktop interface and local workflow for using compatible text-to-speech models.

Citation

@misc{supertonic-mlx,
  title  = {supertonic-mlx: Apple MLX port of Supertonic 3},
  author = {ailuntx},
  year   = {2026},
  url    = {https://github.com/ailuntx/supertonic-mlx},
}

@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},
}