Directory	Quantization	Total size
`models/coreml/`	fp32 (all 4 graphs)	~197 MB
`models/coreml_int8_both/`	fp32 `duration_predictor`+`text_encoder`, int8 `vector_estimator`+`vocoder`	~111 MB

Layout

models/
  coreml/                       # fp32 reference
    duration_predictor_mlprogram.mlpackage
    text_encoder_mlprogram.mlpackage
    vector_estimator_mlprogram.mlpackage
    vocoder_mlprogram.mlpackage
  coreml_int8_both/             # production-recommended variant
    ...
resources/
  voice_styles/                 # 10 voice JSONs (M1-M5, F1-F5)
  onnx/
    tts.json
    unicode_indexer.json
LICENSE
README.md
UPSTREAM.md

Pipeline

The four graphs are wired the same as the upstream Supertonic-3 ONNX: text_processor → duration_predictor → text_encoder → vector_estimator (denoising loop) → vocoder.

Use resources/voice_styles/<voice>.json to load style_dp (8x16) and style_ttl (50x256).

resources/onnx/tts.json carries the model config (sample rate, base chunk size, chunk compress factor, latent dim) needed at runtime.

resources/onnx/unicode_indexer.json holds the multilingual tokenization tables for the v3 Unicode processor.

Quantization details

Method: coremltools.optimize.coreml.linear_quantize_weights
Mode: linear_symmetric
Bits: 8
Granularity: per_channel
Weight threshold: 2048
Applied to: vector_estimator, vocoder
Skipped: weights containing inf/-inf are passed through unquantized (small subset)

Sizes after quantization:

vector_estimator: 127.3 MB → 65.3 MB
vocoder: 48.3 MB → 24.3 MB

Acknowledgments

Original ONNX models, voice styles, and config: Supertone/supertonic-3
CoreML conversion pipeline: adapted from Nooder/supertonic-2-coreml

See UPSTREAM.md for full provenance.

wakirin/supertonic-3-coreml

Supertonic-3 CoreML

Variants

Layout

Pipeline

Quantization details

Acknowledgments

wakirin/
supertonic-3-coreml