Irodori I/C common model artifacts

This distribution contains general model and tokenizer artifacts only.

I: Aratako/Irodori-TTS-v4.1-Small
Revision: 2b28324dc263ed5e6638b3cf3dd94c82ead07b4b
C: Aratako/Irodori-TTS-v4.1-Small-MF
Revision: ccc78f5d480b6e51b69b2d5042a14c4da04fea6e
Copyright (c) 2026 Aratako. MIT License; see Irodori-MIT.txt.
Numerical source: Aratako/Irodori-TTS revision
89f9d8fbd4d51ea019867ee1197725ede1df13c5.

Text encoder and tokenizer ancestry: sbintuitions/modernbert-ja-310m,
text-encoder revision 77675fc96a7e445e982e2ba90246b816efc74ec6.
Copyright (c) 2025 SB Intuitions. MIT; see ModernBERT-MIT.txt.

Audio decoder: Aratako/Semantic-DACVAE-Japanese-32dim,
revision 47376ee24834d7a05a48ebabfe3cde29b3c5e214; MIT as declared
by its model card. Derived from facebook/dacvae-watermarked.
DACVAE: Copyright (c) Meta Platforms, Inc. and affiliates.
All Rights Reserved. Apache License, Version 2.0;
see DACVAE-Apache-2.txt.
Descript Audio Codec ancestry: Copyright (c) 2023-present, Descript.
MIT License; see Descript-MIT.txt.
AudioSeal architecture ancestry: Copyright (c) Meta Platforms,
Inc. and affiliates. MIT License; see AudioSeal-MIT.txt.
No separate AudioSeal checkpoint is included.

MODIFIED FILES: every distributed text/duration/DiT graph and weight
file is a custom ONNX/Q4 conversion of the pinned I/C model above.
The distributed codec graph and tensor data were converted to ONNX
and changed from ConvTranspose to subpixel/polyphase convolution.
Activations and codec are FP32. External tensor data and all graphs
are split into SHA256-addressed transport pieces. These files are
third-party conversions, not upstream releases. ARTIFACTS.json maps
every logical file, content digest and transport piece to these notices.
The Japanese codec's watermark path is bypassed as instructed by its
model card. Synthetic outputs have no watermark guarantee.

Conversion/numerical source attribution:
ngc-shj/irodori-tts-webgpu and the reviewed CharaDock numeric subtree:
Copyright (c) 2026 NOGUCHI Shoji. MIT; see NOGUCHI-Shoji-MIT.txt.
utyujinn/irodori-onnx: Copyright (c) 2026 utyujinn.
MIT License; see utyujinn-MIT.txt.

Use conditions from the Irodori model cards: obtain explicit consent
before cloning or impersonating anyone's voice; do not generate
misleading synthetic speech or misinformation; comply with applicable
law. Generated voices may coincidentally resemble real individuals.
No upstream author endorsement is implied.

The original DACVAE model README retains a conflicting SAM-license
sentence. An AI at Meta maintainer explicitly confirmed on 2025-12-19
that the model weights are Apache-2.0, matching the metadata and the
DACVAE implementation LICENSE. See the official clarification:
https://huggingface.co/facebook/dacvae-watermarked/discussions/1
The stale sentence is documented in SOURCES.md; it is not an additional
noncommercial or SAM restriction asserted by this distribution.

Source links and exact revisions: SOURCES.md.
Full license texts accompany this notice and retain their own terms.