Have a ComfyUI workflow? Get a readable Workflow Doc in your browser.Make a workflow readable.Try it now ↗
First-party CoreML export of Supertonic-3's four non-autoregressive flow-matching graphs, for on-device iOS / Apple Neural Engine. Built by our own pipeline (speech-models/stmodels): weights lifted from the...
Model source
Source description
First-party CoreML export of Supertonic-3's four non-autoregressive flow-matching graphs, for
on-device iOS / Apple Neural Engine. Built by our own pipeline (speech-models/stmodels): weights
lifted from the ONNX
initializers → PyTorch → (, FP32, iOS18+).
Sources
1 sourceVerified Aug 29
Model artifacts
4 artifactsSource excerpts
2 excerpts
nn.Modulecoremltoolsmlprogram| Module | mlpackage | parity max|Δ| |
|---|---|---|
| Duration predictor | DurationPredictor.mlpackage | 7.2e-06 ✓ |
| Vector estimator (ODE denoiser) | VectorEstimator.mlpackage | 2.5e-03 ✓ |
| Vocoder | Vocoder.mlpackage | 3.0e-04 ✓ |
| Text encoder | TextEncoder.mlpackage | mean 2.5e-04 (max 2.5e-2 at isolated positions) |
Text/duration use fixed T=128 (relpos attention has T-dependent pad widths — pad/segment text to
128); vocoder + vector-estimator use a dynamic latent-length RangeDim. The host runs the flow-matching
ODE loop (vector_estimator ×total_steps) — the graphs contain no control flow. Assets to drive them:
tts.json, unicode_indexer.json (G2P-free tokenizer table), voice_styles/*.json.
FP32 = parity reference. For ANE residency, use the mixed-precision
Supertonic-3-CoreML-FP16— vocoder + duration FP16, text-encoder + vector-estimator FP32; measured transparent at 47–51 dB mag-STFT SNR.
Supertone/supertonic-3
(commit 3cadd1ee6394adea1bd021217a0e650ede09a323), Supertone Inc., arXiv:2503.23108 — OpenRAIL-M
(use-based restrictions carry over: no non-consensual impersonation/deepfakes, etc.).TTSInterface CoreML model.VectorEstimator.mlpackage/Data/com.apple.CoreML/weights/weight.bin
bin · 243 MB · SHA-256 593cb7df14ef…e1fd · Hugging Face
DownloadVocoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
bin · 96.6 MB · SHA-256 f0f325c9e90f…ac6d · Hugging Face
Download--- license: openrail base_model: Supertone/supertonic-3 tags: - text-to-speech - tts - coreml - on-device - ios - ane - supertonic - flow-matching library_name: coreml pipeline_tag: text-to-speech --- # SupertonicTTS-3 — CoreML (.mlpackage, iOS) **First-party CoreML export** of Supertonic-3's four non-autoregressive flow-matching graphs, for on-device iOS / Apple Neural Engine. Built by our own pipeline (`speech-models/stmodels`): weights lifted from the [`Supertone/supertonic-3`](https://huggingface.co/Supertone/supertonic-3) ONNX initializers → PyTorch `nn.Module` → `coremltools` (`mlprogram`, FP32, iOS18+). ## Graphs & parity (FP32, vs ONNX Runtime) | Module | mlpackage | parity max\|Δ\| | |---|---|---| | Duration predictor | `DurationPredictor.mlpackage` | 7.2e-06 ✓ | | Vector estimator (ODE denoiser) | `VectorEstimator.mlpackage` | 2.5e-03 ✓ | | Vocoder | `Vocoder.mlpackage` | 3.0e-04 ✓ | | Text encoder | `TextEncoder.mlpackage` | **mean 2.5e-04** (max 2.5e-2 at isolated positions) | Text/duration use **fixed T=128** (relpos attention has T-dependent pad widths — pad/segment text to 128); vocoder + vector-estimator use a dynamic latent-length `RangeDim`. The host runs the flow-matching ODE loop (`vector_estimator` ×`total_steps`) — the graphs contain no control flow. Assets to drive them: `tts.json`, `unicode_indexer.json` (G2P-free tokenizer table), `voice_styles/*.json`. > FP32 = parity reference. For **ANE residency**, use the mixed-precision > [`Supertonic-3-CoreML-FP16`](https://huggingface.co/aufklarer/Supertonic-3-CoreML-FP16) — vocoder + duration FP16, > text-encoder + vector-estimator FP32; measured transparent at **47–51 dB** mag-STFT SNR. ## Attribution & license - **Weights**: derivative of [`Supertone/supertonic-3`](https://huggingface.co/Supertone/supertonic-3) (commit `3cadd1ee6394adea1bd021217a0e650ede09a323`), Supertone Inc., [arXiv:2503.23108](https://arxiv.org/abs/2503.23108) — **OpenRAIL-M** (use-based restrictions carry over: no non-consensual impersonation/deepfakes, etc.). ## Other Supertonic-3 formats - [**Supertonic-3 — CoreML (FP16)**](https://huggingface.co/aufklarer/Supertonic-3-CoreML-FP16) — mixed-precision ANE variant (47–51 dB). - [**Supertonic-3 — ONNX (INT8)**](https://huggingface.co/soniqo/Supertonic-3-ONNX-INT8) — server / desktop (ONNX Runtime). - [**Supertonic-3 — LiteRT**](https://huggingface.co/soni...
Source context: 2860 downloads · 0 likes · Pipeline text-to-speech · Library coreml · Repo aufklarer/Supertonic-3-CoreML