# Common model sources and licensing

Reviewed 2026-10-01. This directory accompanies only the 924,139,872 bytes of model/tokenizer content identified in ARTIFACTS.json. ONNX Runtime, JS application code and private voice-derived files are outside this public asset distribution.

- **I**: [Aratako/Irodori-TTS-v4.1-Small](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small), revision `2b28324dc263ed5e6638b3cf3dd94c82ead07b4b`; MIT and model-card use restrictions. Numerical implementation: [Aratako/Irodori-TTS at the recorded pin](https://github.com/Aratako/Irodori-TTS/tree/89f9d8fbd4d51ea019867ee1197725ede1df13c5).
- **C**: [Aratako/Irodori-TTS-v4.1-Small-MF](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small-MF), revision `ccc78f5d480b6e51b69b2d5042a14c4da04fea6e`; MIT and model-card use restrictions.
- **Text backbone/tokenizer**: [sbintuitions/modernbert-ja-310m license](https://huggingface.co/sbintuitions/modernbert-ja-310m/blob/77675fc96a7e445e982e2ba90246b816efc74ec6/LICENSE), text-encoder revision `77675fc96a7e445e982e2ba90246b816efc74ec6`; MIT, SB Intuitions 2025. The copied tokenizer is additionally pinned by its SHA256 in ARTIFACTS.json.
- **Codec**: [Aratako/Semantic-DACVAE-Japanese-32dim](https://huggingface.co/Aratako/Semantic-DACVAE-Japanese-32dim), revision `47376ee24834d7a05a48ebabfe3cde29b3c5e214`; declared MIT. Its card identifies `facebook/dacvae-watermarked` ancestry. The fine-tune's exact original base-weight commit is not stated; the pinned Japanese checkpoint and exported artifact hashes are stated.
- **DACVAE ancestry**: [Meta DACVAE Apache-2.0 LICENSE](https://github.com/facebookresearch/dacvae/blob/main/LICENSE), [base model](https://huggingface.co/facebook/dacvae-watermarked), [official model-weight license clarification](https://huggingface.co/facebook/dacvae-watermarked/discussions/1).
- **Descript ancestry**: [Descript Audio Codec MIT LICENSE](https://github.com/descriptinc/descript-audio-codec/blob/main/LICENSE), Copyright (c) 2023-present, Descript.
- **AudioSeal ancestry**: [official project](https://github.com/facebookresearch/audioseal), [Meta MIT LICENSE](https://github.com/facebookresearch/audioseal/blob/main/LICENSE). This notice documents DACVAE's adapted architecture. No separate AudioSeal checkpoint is bundled.
- **ngc-shj exporter**: [source](https://github.com/ngc-shj/irodori-tts-webgpu/tree/aa3b6390018bb09a2e461c95d1f55992c06e197d), MIT, NOGUCHI Shoji 2026.
- **CharaDock numeric subtree**: [source](https://github.com/ochisamu/CharaDock/tree/c5e8ab2493050dd7602cbe3606eb0a425b5e1a96), reviewed numeric portion MIT, NOGUCHI Shoji 2026. The entire application is not included.
- **utyujinn exporter**: [source](https://github.com/utyujinn/irodori-onnx/tree/386eaf48a3b482b8f8bd509b6c65b904a0a512df), MIT, utyujinn 2026.

## Base-codec documentation mismatch

In `facebook/dacvae-watermarked/README.md` at revision `8680102d141858a21bd533543966a2eb2e569f92`, the metadata is Apache-2.0 while the prose says “This project is licensed under the SAM License”. [Pinned README](https://huggingface.co/facebook/dacvae-watermarked/blob/8680102d141858a21bd533543966a2eb2e569f92/README.md).

In the official discussion, **lematt1991**, identified as **AI at Meta org**, answered the explicit model-weight licensing question on 2025-12-19: “Yes, also apache-2.0. Updated the model card”. This primary-source statement and the Apache metadata resolve the license identity used here. [Clarification](https://huggingface.co/facebook/dacvae-watermarked/discussions/1). The [README cleanup PR](https://huggingface.co/facebook/dacvae-watermarked/discussions/2) remains open. The leftover sentence should not be presented as an unconfirmed noncommercial condition or a redistribution ban.

The Meta README revision is license-evidence provenance, not an assertion of the original fine-tuning checkpoint revision.