README.md

metadata

license: openrail
pipeline_tag: text-to-speech
base_model: Supertone/supertonic-3
tags:
  - text-to-speech
  - tts
  - mlx
  - supertonic

supertonic-3-mlx

Native-MLX conversion of Supertone/supertonic-3 (99M-parameter flow-matching TTS, 44.1 kHz, 10 preset voices, 32 languages) for use with mlx-audio-swift's SupertonicModel (model_type supertonic).

Contents

• config.json — model_type supertonic, sample_rate 44100, hop 512, latent_dim 24, chunk_factor 6, cfg_scale 4.0, default_steps 8
• duration_predictor.safetensors, text_encoder.safetensors, vector_estimator.safetensors, vocoder.safetensors — the four ONNX sub-graphs' initializers, converted to MLX layout (Conv [O,I,K] → [O,K,I])
• unicode_indexer.json — 65,536-entry BMP codepoint → embedding-row table
• voice_styles/{M1..M5,F1..F5}.json — preset style vectors (style_ttl [1,50,256], style_dp [1,8,16])

Conversion pipeline

ONNX initializers were extracted with onnx.numpy_helper, Conv weights transposed to MLX conv1d layout, keys renamed to stable dotted paths, and the result validated stage-by-stage against ONNX Runtime at ≤1e-4 max-abs-error (end-to-end ≤1e-3 on identical injected noise; ~69 dB SNR, perceptually transparent). CFG and the Euler step baked into vector_estimator.onnx are factored out and applied by the runtime.

License

BigScience Open RAIL-M with use-based restrictions — see LICENSE and NOTICE. These are format-converted (modified) files of the original Supertone release; all credit for the model belongs to Supertone Inc.