lmh8602/supertonic-3-coreml-t320

Supertonic-3 Core ML — text window T=320

A Core ML re-conversion of Supertone/supertonic-3 with the text encoder's fixed text length set to T=320 instead of the T=128 used by the published Core ML conversions.

This is a derivative of the model under the BigScience Open RAIL-M license. Use of this repository is subject to the use-based restrictions in Attachment A of the included LICENSE (Open RAIL-M §5). By downloading or using these files you agree to those restrictions. See "License" below.

Why T=320

Supertonic-3's text window is measured in NFKD Unicode scalars, not characters. Hangul syllables decompose into 2–3 jamo each, so Korean text costs roughly 2.1–2.6 scalars per character and T=128 fills up at about 57 Korean characters. Longer chunks are silently truncated at the tail. The upstream ONNX graph leaves text_length dynamic — 128 is a property of the conversion, not of the model — so the text-dependent parts were re-exported with T=320, which holds about 145 Korean characters.

Contents

Path	Note
TextEncoder.mlmodelc	re-converted, T=320, fp32
DurationPredictor.mlmodelc	re-converted, T=320, fp32
VectorEstimatorVariants/VectorEstimator_L{128,256,512}_int8.mlmodelc	re-converted, T=320, int8 ANE latent buckets
Vocoder.mlmodelc	unchanged from the T=128 conversion (no T axis)
tts.json, unicode_indexer.json, voice_styles/	unchanged upstream companion files
manifest.txt	one repo-relative path per line — the file list a plain-HTTPS mirror client needs
LICENSE	BigScience Open RAIL-M, as distributed with the upstream model

Modified files are marked re-converted above (Open RAIL-M §4.c). Everything else is carried over from upstream unchanged.

Loading

The layout matches what FluidAudio expects under <cache root>/supertonic-3/. Stock FluidAudio assumes T=128; a build that reads the pinned T off the loaded text_encoder description is needed to use this set correctly — see the probe/t320 branch of lmh8602/FluidAudio, which also accepts a plain-HTTPS mirror URL and reads manifest.txt from it. This repository can be used as such a mirror via its resolve/main/ URLs.

Conversion

Exported from the upstream ONNX with coremltools; precision choices match the published T=128 set (fp32 text encoder/duration predictor, int8 vector-estimator buckets) so that the only axis that differs is T. Relative error against the ONNX reference is in the same range as the published T=128 conversion measured the same way.

License

• Upstream model: Supertone/supertonic-3, BigScience Open RAIL-M (see LICENSE).
• This derivative is distributed under the same license. The use-based restrictions of Attachment A apply to any use or redistribution of these files and must be passed on to anyone you distribute them to (§4.a). Keep the LICENSE file with the model (§4.b) and retain the notices it contains (§4.d).
• Conversion work by lmh8602 (2026). No warranty; see §7–§8 of the license.