THIRD_PARTY_NOTICES.md

# Third-party notices

## Irodori-TTS

- v4.1 source: https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small (revision `2b28324dc263ed5e6638b3cf3dd94c82ead07b4b`)
- v4 implementation: https://github.com/Aratako/Irodori-TTS (revision `8224dafb46d0aba89209a8f905f1cb7e3299d9c1`); shared backbone export and tensor-based attention mask handling for dynamic ONNX inputs.
- Copyright: Chihiro Arata and contributors
- Sources: https://huggingface.co/Aratako/Irodori-TTS-500M-v3 and https://huggingface.co/Aratako/Irodori-TTS-600M-v3-VoiceDesign
- License: MIT (`LICENSES/MIT.txt`)
- Changes: converted from PyTorch checkpoints to ONNX; split static conditioning projection into `context_kv.onnx`; fixed export dimensions for the iOS runtime.

The responsible-use restrictions and disclaimers from the upstream model cards are reproduced in this repository's README.

## Semantic-DACVAE-Japanese-32dim

- Copyright: Chihiro Arata and contributors
- Source: https://huggingface.co/Aratako/Semantic-DACVAE-Japanese-32dim
- License: MIT (`LICENSES/MIT.txt`)
- Changes: converted encoder and decoder to ONNX for the iOS runtime.

## llm-jp-3-150m tokenizer

- Copyright: LLM-jp / National Institute of Informatics contributors
- Source: https://huggingface.co/llm-jp/llm-jp-3-150m
- License: Apache License 2.0 (`LICENSES/Apache-2.0.txt`)
- Files redistributed: tokenizer JSON and tokenizer configuration metadata.

## ONNX export code

- Source: https://github.com/Aratako/Irodori-TTS
- License: MIT (`LICENSES/MIT.txt`)

## ModernBERT-ja-310m (v4.1 encoder and tokenizer)

- Source: https://huggingface.co/sbintuitions/modernbert-ja-310m
- Copyright: SB Intuitions and contributors
- License: MIT
- The fine-tuned encoder and exact tokenizer are redistributed as part of the Irodori-TTS-v4.1-Small conversion. Text and caption share one encoder graph.

## DACVAE / Descript Audio Codec ancestry (ONNX and Core ML)

- Semantic-DACVAE-Japanese-32dim source revision: `47376ee24834d7a05a48ebabfe3cde29b3c5e214` (MIT as declared by its model card).
- Base weights: https://huggingface.co/facebook/dacvae-watermarked (model metadata: Apache-2.0).
- DACVAE implementation: https://github.com/facebookresearch/dacvae (Apache License 2.0; `LICENSES/Apache-2.0.txt`). Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved.
- DAC architecture: https://github.com/descriptinc/descript-audio-codec (MIT; `LICENSES/Descript-MIT.txt`). Copyright (c) 2023-present, Descript.
- The base model README also contains a conflicting “SAM License” sentence; the linked DACVAE implementation LICENSE is Apache-2.0. This distribution preserves the Apache-2.0 license and does not relicense upstream components as MIT.
- Changes by the Onsei conversion: exported the decoder to a float32 Core ML ML Program using coremltools 9.0, folded weight normalization, froze constant expressions, and bounded batch/time input dimensions. No quantization or retraining. Watermarking remains bypassed as instructed by the Japanese model card. The converted model is not a watermarked-audio guarantee.
- `native_decoder/manifest.json` records the source ONNX digest, converted file digests and numerical verification. Core ML conversion does not grant any additional rights to upstream weights.