Irodori Tts Voicedesign GGUF

Irodori-TTS-600M-v3-VoiceDesign — GGUF

GGUF conversions of Aratako/Irodori-TTS-600M-v3-VoiceDesign for the CrispASR irodori-tts backend.

Irodori-TTS VoiceDesign is a rectified-flow diffusion transformer (RF-DiT) for Japanese text-to-speech with three conditioning modes:

Mode	Inputs	Use case
Pure Voice Design	Text + Caption	Generate diverse voices from text descriptions alone
Style-Controlled Cloning	Text + Caption + Ref audio	Clone a voice while controlling emotion/delivery
Unconditional	Text only	Neutral synthesis (no caption, no reference)

Architecture

• TextEncoder: 10-layer RoPE transformer (512-dim, 8 heads, llm-jp-3-150m tokenizer)
• CaptionEncoder: 10-layer RoPE transformer (same architecture as text encoder) — encodes style/emotion descriptions
• ReferenceLatentEncoder: 8-layer transformer (768-dim, 12 heads) — encodes speaker identity from DAC-VAE latents
• DiT: 12 DiffusionBlock layers (1280-dim, 20 heads) with LowRankAdaLN + JointAttention (text + speaker + caption KV) + half-RoPE + SwiGLU
• DurationPredictor: token_sum_dual_adarn_zero — speaker + caption dual modulation
• DAC-VAE decoder: Semantic-DACVAE-Japanese-32dim (48 kHz, 32-dim latent)
• ODE solver: Euler RF with independent text/speaker/caption CFG

Files

File	Size	Description
irodori-tts-600m-v3-voicedesign-q4_k.gguf	525 MB	Q4_K quantized (recommended)
irodori-tts-600m-v3-voicedesign-f16.gguf	1.2 GB	F16 full precision
dacvae-ja-32dim-f16.gguf	135 MB	DAC-VAE codec (required companion)

Usage with CrispASR

# Voice Design — generate a voice from a text description:
crispasr --backend irodori-tts \
    -m irodori-tts-600m-v3-voicedesign-q4_k.gguf \
    --codec-model dacvae-ja-32dim-f16.gguf \
    --instruct "落ち着いた大人の男性。フォーマルな場で、深く響く声で丁寧に話している。" \
    --tts "こんにちは、世界。" \
    --tts-output voicedesign.wav

# Style-controlled voice cloning:
crispasr --backend irodori-tts \
    -m irodori-tts-600m-v3-voicedesign-q4_k.gguf \
    --codec-model dacvae-ja-32dim-f16.gguf \
    --voice reference.wav --i-have-rights \
    --instruct "元気で明るい女性。笑顔で楽しそうに話している。" \
    --tts "今日はいい天気ですね!" \
    --tts-output cloned_styled.wav

# Auto-download (fetches both model + codec):
crispasr --backend irodori-tts-voicedesign -m auto --auto-download \
    --instruct "優しい声の女性。" \
    --tts "テスト。" --tts-output out.wav

Environment variables

Variable	Default	Description
CRISPASR_IRODORI_CAPTION	—	Caption text (alternative to --instruct)
CRISPASR_IRODORI_CFG_TEXT	3.0	Text CFG guidance scale
CRISPASR_IRODORI_CFG_SPEAKER	5.0	Speaker CFG guidance scale
CRISPASR_IRODORI_CFG_CAPTION	3.0	Caption CFG guidance scale
CRISPASR_IRODORI_ODE_STEPS	40	Number of Euler ODE steps

Parity

Diff harness results against the original PyTorch implementation:

Stage	F16 vs Python	Q4_K vs Python
text_state	cos=1.000000, max_abs=3.96e-04	—
v_pred_step0	cos=1.000000, max_abs=8.46e-04	cos=0.996491, max_abs=4.66e-01

License

MIT (same as the original model). See Aratako/Irodori-TTS-600M-v3-VoiceDesign for usage restrictions (no impersonation without consent, AI-generated audio disclaimer required for caption-only outputs).

Credits

• Original model: Aratako
• GGUF conversion & runtime: CrispASR

Provenance and EU AI Act Art. 53 note

• Upstream model: Aratako/Irodori-TTS-600M-v3-VoiceDesign — published by Aratako.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.