| --- |
| license: openrail |
| language: |
| - en |
| library_name: coreml |
| pipeline_tag: text-to-speech |
| tags: |
| - tts |
| - text-to-speech |
| - coreml |
| - mlprogram |
| - supertonic |
| - ios |
| - macos |
| - on-device |
| --- |
| |
| # Supertonic v1 Core ML |
|
|
| Core ML ML Program exports of **Supertonic v1 English** for iOS and macOS. |
|
|
| This repository hosts the heavy binary artifacts for the |
| [`eborges78/supertonic-1-coreml`](https://github.com/eborges78/supertonic-1-coreml) |
| GitHub project. The GitHub repo holds the code, conversion scripts, runtime |
| contract, Swift package, tests, and CI. This Hugging Face repo holds the |
| weights. |
|
|
| ## What Is Included |
|
|
| Core ML packages (fixed-shape, ML Program, iOS 16+): |
|
|
| - `models/supertonic-v1/coreml/duration_predictor_mlprogram.mlpackage` |
| - `models/supertonic-v1/coreml/text_encoder_mlprogram.mlpackage` |
| - `models/supertonic-v1/coreml/vector_estimator_mlprogram.mlpackage` |
| - `models/supertonic-v1/coreml/vocoder_mlprogram.mlpackage` |
|
|
| Source ONNX models (for reconverting or inspection): |
|
|
| - `resources/supertonic-v1/onnx/duration_predictor.onnx` |
| - `resources/supertonic-v1/onnx/text_encoder.onnx` |
| - `resources/supertonic-v1/onnx/vector_estimator.onnx` |
| - `resources/supertonic-v1/onnx/vocoder.onnx` |
|
|
| The small runtime files (voice styles, embedding tables, `tts.json`, |
| `unicode_indexer.bin`) live in the GitHub repository. |
|
|
| ## Export Configuration |
|
|
| ```text |
| format=mlprogram |
| minimum_deployment_target=iOS16 |
| max_text_len=300 |
| max_seconds=20 |
| latent_length=288 |
| sample_rate=44100 |
| ``` |
|
|
| ## Quick Download |
|
|
| From Python, using the `huggingface_hub` CLI: |
|
|
| ```sh |
| pip install -U huggingface_hub |
| huggingface-cli download eborges78/supertonic-v1-coreml \ |
| --local-dir ./supertonic-v1-coreml |
| ``` |
|
|
| Or use the helper script in the GitHub repo: |
|
|
| ```sh |
| git clone https://github.com/eborges78/supertonic-1-coreml.git |
| cd supertonic-1-coreml |
| python scripts/download_models.py |
| ``` |
|
|
| ## Runtime Contract |
|
|
| The four packages are fixed-shape and expect externalized text embeddings. |
| See [`docs/runtime-contract.md`](https://github.com/eborges78/supertonic-1-coreml/blob/main/docs/runtime-contract.md) |
| in the GitHub repo for the exact tensor shapes and the integration steps. |
|
|
| A minimal Swift integration reference is shipped as a SwiftPM library: |
| [`Sources/SupertonicV1CoreML/SupertonicV1CoreMLPipeline.swift`](https://github.com/eborges78/supertonic-1-coreml/blob/main/Sources/SupertonicV1CoreML/SupertonicV1CoreMLPipeline.swift). |
|
|
| ## Provenance and License |
|
|
| - Upstream model: [`Supertone/supertonic`](https://huggingface.co/Supertone/supertonic) — OpenRAIL-M. |
| - Upstream project: [`supertone-inc/supertonic`](https://github.com/supertone-inc/supertonic). |
| - Core ML conversion reference: [`Nooder/supertonic-2-coreml`](https://github.com/Nooder/supertonic-2-coreml). |
|
|
| The weights in this repository are redistributed under the upstream OpenRAIL-M |
| license. Use-based restrictions from the upstream license are preserved. See |
| the [GitHub repo's `MODEL_LICENSE.md`](https://github.com/eborges78/supertonic-1-coreml/blob/main/MODEL_LICENSE.md) |
| for details. |
|
|
| The conversion scripts and sample integration code on GitHub are MIT-licensed. |
|
|
| ## Acknowledgements |
|
|
| - Supertone for releasing Supertonic v1. |
| - The Supertonic 2 Core ML community work that informed parts of this |
| conversion approach. |
|
|
