| # OmniVoice Core ML bundles (on-device iOS client export) | |
| Private artifact drop for **on-device** experiments. Upstream weights: [`k2-fsa/OmniVoice`](https://huggingface.co/k2-fsa/OmniVoice). | |
| ## Contents | |
| | Path | Description | | |
| |------|-------------| | |
| | `OmniVoice_Vocoder.mlpackage` | Higgs **decode** (discrete codes → waveform). | | |
| | `OmniVoice_GeneratorStep.mlpackage` | **Real** `OmniVoice.forward` logits (`torch.export` + coremltools). | | |
| | `OmniVoice_AudioEncoder.mlpackage` | **Real** Higgs **encode** (waveform → codes). | | |
| | `OmniVoice_LM.mlpackage` | **Stub** I/O only (not trained weights). | | |
| | `OmniVoice_SpeakerEncoder.mlpackage` | **Stub** placeholder. | | |
| | `tokens.txt` | Text tokenizer vocabulary for Swift. | | |
| Export scripts live in the local project under `Scripts/export_omnivoice_coreml/`. See `Docs/OmniVoice_CoreML_Contract.md` for Swift vs Python I/O notes. | |
| **License:** Follow the license terms of `k2-fsa/OmniVoice` and third-party components (Qwen3, Higgs, etc.). Do not redistribute if your use case is not permitted by those licenses. | |