
Format	JSON graph topology + NPZ initializers
Runtime	`ailuntx/supertonic-mlx`
Official code	`supertone-inc/supertonic`
Sample rate	44.1 kHz
HF Space	`mlx-community/supertonic-3`
Hardware	Runs on HF Linux CPU fallback; Apple Silicon recommended locally

Quick Start

bash

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

Layout

text

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

Conversion Notes

Component	Source	MLX handling
ONNX graphs	`Supertone/supertonic-3`	graph topology exported to JSON
initializers	official ONNX assets	saved as NPZ arrays
runtime ops	Supertonic ONNX subset	implemented in `ailuntx/supertonic-mlx` with MLX arrays

Validation

The MLX graph runtime has been checked against ONNX Runtime on the official assets; per-stage maximum absolute errors are around 1e-5. The HF Space API has returned audio successfully with real wall-time status reporting.

License

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

Citation

bibtex

@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},
}