Irodori-TTS-MLX-DACVAE-Codec

Converted Semantic-DACVAE MLX codec artifact for t0yohei/Irodori-TTS-MLX.

This repository is a hosted codec companion artifact. It is separate from the hosted RF-DiT weights repositories and contains only the converted DACVAE codec artifact plus manifest, checksums, provenance, and validation reports.

Contents

  • irodori_dacvae_codec_manifest.json
  • dacvae-codec.npz
  • codec_metadata.json
  • checksums.sha256
  • validation/convert-dacvae-codec.json
  • validation/dacvae-decode-parity.json
  • validation/dacvae-encode-parity.json

Provenance

  • Upstream codec: Aratako/Semantic-DACVAE-Japanese-32dim
  • Upstream source file: weights.pth
  • Upstream revision: 47376ee24834d7a05a48ebabfe3cde29b3c5e214
  • Converter: t0yohei/Irodori-TTS-MLX at 19ae117b2b553c5672d7ee3caa47bf7de60397b2
  • DACVAE package revision used for interpretation: 414c20785fc3a28373073ea8ef7a1316eeeaca6e
  • License review: approved for personal OSS, research, and development publication; see https://github.com/t0yohei/Irodori-TTS-MLX/issues/188

This is a converted derivative artifact. It is not the original upstream weights.pth.

Runtime Capabilities

  • supports_mlx_decode: true
  • supports_mlx_encode: true
  • requires_pytorch_fallback: false
  • sample rate: 48000
  • hop length: 1920
  • latent dim: 32

Validation

Decode parity against the upstream PyTorch bridge passed:

  • max_abs: 3.0517578125e-05
  • mean_abs: 1.4040205087439972e-07
  • rmse: 2.0699590095318854e-06
  • cosine: 0.9999999403953552

Encode parity against the upstream PyTorch bridge passed:

  • max_abs: 1.33514404296875e-05
  • mean_abs: 2.287564711878076e-06
  • rmse: 3.305609425297007e-06
  • cosine: 1.0

Usage

Pin the codec artifact revision for reproducible runs. The first published artifact revision is bb89840af0deb729cc7a8e4ba5ebddb49e2b3e78.

No-reference v3 generation can run the codec fully through MLX:

irodori-tts-generate \
  --weights-repo t0yohei/Irodori-TTS-MLX-500M-v3 \
  --weights-revision 078ffb11ffad92e6dde237a6abef730f4341b359 \
  --codec-runtime-mode mlx \
  --codec-artifact-repo t0yohei/Irodori-TTS-MLX-DACVAE-Codec \
  --codec-artifact-revision bb89840af0deb729cc7a8e4ba5ebddb49e2b3e78 \
  --text "こんにちは。今日は良い天気です。" \
  --no-reference \
  --output /tmp/irodori-v3-mlx-codec.wav

For reference-audio generation, use the same codec artifact with --codec-runtime-mode mlx so both reference encode and waveform decode use MLX.

License

The hosted artifact is published under MIT to match the upstream model card. Preserve upstream attribution and follow applicable laws and rights for prompts, reference audio, and generated audio.

Acknowledgements:

  • Aratako/Semantic-DACVAE-Japanese-32dim
  • Aratako/Semantic-DACVAE-Japanese
  • facebook/dacvae-watermarked
  • facebookresearch/dacvae
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for t0yohei/Irodori-TTS-MLX-DACVAE-Codec