Converted MLX RF-DiT weights for Aratako/Irodori-TTS-500M-v3, produced for the unofficial t0yohei/Irodori-TTS-MLX runtime.

This is a converted MLX artifact, not the original upstream checkpoint. It is published as an approved hosted RF-DiT weights layout for the v0.2 irodori-tts-mlx runtime contract.

This repository is gated. Accept the repository conditions on Hugging Face and use an authenticated huggingface_hub session before resolving it with --weights-repo.

PYTHONPATH=/path/to/Irodori-TTS:${PYTHONPATH:-} \
irodori-tts-generate \
  --weights-repo t0yohei/Irodori-TTS-MLX-500M-v3 \
  --weights-revision 078ffb11ffad92e6dde237a6abef730f4341b359 \
  --text "こんにちは。今日は良い天気です。" \
  --no-reference \
  --output /tmp/irodori-v3-hosted.wav \
  --preset balanced

This v3 artifact supports predicted-duration generation. Omit --seconds to exercise the v3 duration predictor.

By default, the current irodori-tts-mlx runtime resolves the separate approved MLX DACVAE codec artifact for encode/decode. RF-DiT weights and DACVAE codec artifacts intentionally stay in separate repositories. Pin the codec contract with --codec-runtime-mode mlx, --codec-artifact-repo t0yohei/Irodori-TTS-MLX-DACVAE-Codec, and --codec-artifact-revision bb89840af0deb729cc7a8e4ba5ebddb49e2b3e78 when reproducibility matters.

This repository includes the converted RF-DiT/text/speaker/duration-predictor weights and lightweight loader metadata for irodori-tts-mlx. It does not bundle Semantic-DACVAE codec weights, upstream source code, reference audio, generated audio samples, Hugging Face cache snapshots, training data, tokenizer assets, or unaudited model artifacts. The included tokenizer_config.json is metadata for the runtime contract, not a redistributed tokenizer asset.

The current public generation path uses the separate approved MLX DACVAE codec artifact instead of the removed upstream PyTorch DACVAE bridge. This repository does not bundle that codec artifact; resolve it with the runtime default, --codec-artifact-repo, or a local --codec-artifact-dir.

The upstream model card references SilentCipher watermarking in generated outputs. This converted RF-DiT artifact does not implement, redistribute, or bundle SilentCipher itself.

See the redistribution audit: https://github.com/t0yohei/Irodori-TTS-MLX/blob/main/docs/preconverted_weights_redistribution_audit.md

For the complete hosted-layout contract and local conversion fallback, see: https://github.com/t0yohei/Irodori-TTS-MLX/blob/main/docs/hosted_weights_usage.md