README.md

---
license: other
language:
  - ja
pipeline_tag: text-to-speech
tags:
  - coreml
  - onnx
  - ios
  - on-device
  - irodori-tts
---

# Onsei iOS ONNX and Core ML Model Packs

ONNX model packs exported for the Onsei iOS client from the following upstream models:

- [`Aratako/Irodori-TTS-v4.1-Small-MF`](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small-MF), source revision `ccc78f5d480b6e51b69b2d5042a14c4da04fea6e`
- [`Aratako/Irodori-TTS-v4.1-Small`](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small), source revision `2b28324dc263ed5e6638b3cf3dd94c82ead07b4b`
- [`Aratako/Irodori-TTS-500M-v3`](https://huggingface.co/Aratako/Irodori-TTS-500M-v3), source revision `236c1e56591279fc24e3c1bf6609fc06e48dde28`
- [`Aratako/Irodori-TTS-600M-v3-VoiceDesign`](https://huggingface.co/Aratako/Irodori-TTS-600M-v3-VoiceDesign), source revision `e863a3a93e652e09afeff3e84823a206a0a60314`
- [`Aratako/Semantic-DACVAE-Japanese-32dim`](https://huggingface.co/Aratako/Semantic-DACVAE-Japanese-32dim)
- Tokenizer from [`llm-jp/llm-jp-3-150m`](https://huggingface.co/llm-jp/llm-jp-3-150m)

These files are converted runtime artifacts, not new model weights. They are not affiliated with or endorsed by the upstream authors.

## Packs

- `v4.1-small-mf/`: MeanFlow-distilled v4.1 Small. It preserves text, Voice Design, reference-audio cloning and duration conditioning while using one conditional DiT evaluation per step. The app defaults to four steps and ignores RF-only CFG, sway, rescale and speaker-KV controls.
- `v4.1-small/`: unified Japanese TTS, reference cloning and VoiceDesign, with the improved v4.1 duration predictor. `text_encoder.onnx` contains the shared ModernBERT backbone and both `text_state` and `caption_state` outputs. The app reuses one session for both inputs. No separate `caption_encoder.onnx` is needed. Requires the v4.1-compatible iOS client.
- `500m-v3/`: standard Japanese TTS and reference-audio voice cloning.
- `600m-v3-vd/`: VoiceDesign model with caption conditioning.

Each pack contains the split `context_kv.onnx` + `dit_step.onnx` execution path used by the Onsei iOS app, DACVAE encode/decode, speaker and text encoders, duration prediction, configuration, and tokenizer data. `manifest.json` records file sizes and SHA-256 digests used by the app to verify downloads.

## Licenses and attribution

The Irodori-TTS weights, Japanese Semantic-DACVAE adaptations, ModernBERT-ja-310m encoder/tokenizer and conversion code declare MIT. The underlying Meta DACVAE weights/implementation declare Apache License 2.0; Descript DAC uses MIT. The legacy llm-jp tokenizer uses Apache License 2.0. See `LICENSES/` and `THIRD_PARTY_NOTICES.md`. Each component remains subject to its own license; no relicensing is implied.

## Responsible use

The upstream model cards request that users:

1. Do not clone or impersonate a person's voice without explicit consent.
2. Do not create misleading deepfakes or synthetic speech intended to spread misinformation.
3. Recognize that text-only output may coincidentally resemble a real voice.
4. Take responsibility for compliance with applicable law.

The ONNX export used here does not add a SilentCipher watermark. Applications using these files should clearly disclose synthetic audio where appropriate and must enforce consent for voice cloning.

## Citation

```bibtex
@misc{irodori-tts-v3,
  author = {Chihiro Arata},
  title = {Irodori-TTS: A Flow Matching-based Text-to-Speech Model with Emoji-driven Style Control},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {https://huggingface.co/Aratako/Irodori-TTS-500M-v3}
}
```

## Native GPU decoder (ios-v3 and later)

The v4 packs contain the verified float32 Core ML decoder. New clients download it together with the ONNX pack and verify each file using SHA-256. The app measures CPU/GPU speed and numerical parity on the device before selecting the faster decoder. Existing immutable release tags remain unchanged. The sidecar includes licenses, attribution and conversion notices.