cstr/irodori-tts-GGUF

Irodori-TTS-500M-v3 — GGUF (ggml-quantised)

GGUF / ggml conversion of Aratako/Irodori-TTS-500M-v3 for use with CrispStrobe/CrispASR.

Irodori-TTS is a 500M-param Rectified Flow Diffusion Transformer (RF-DiT) text-to-speech model with zero-shot voice cloning via DAC-VAE latent conditioning. Japanese-focused with a 102K-vocab sarashina2.2 tokenizer. 48 kHz mono output.

License: MIT (follows upstream Irodori-TTS license).

Files

File	Quant	Size	Notes
irodori-tts-500m-v3-f16.gguf	F16	~1.9 GB	Reference quality
irodori-tts-500m-v3-q4_k.gguf	Q4_K	~852 MB	Recommended — fits 8 GB RAM
irodori-tts-500m-v3-q8_0.gguf	Q8_0	~896 MB	Near-lossless
irodori-tts-ref.gguf	F32	~4 KB	Reference activations for diff harness

Architecture

Text Input (Japanese / mixed)
  │
sarashina2.2 Tokenize (102K vocab, BPE)
  │
TextEncoder (14L, 1280d, 10 heads, RoPE + SwiGLU)
  │── Each position: self-attention + gated residual + SwiGLU FFN
  │
  ├── [Optional] ReferenceLatentEncoder (14L, 1280d)
  │    └── DAC-VAE latent from reference audio → speaker conditioning
  │
DiT Backbone (24L, 2048d, 16 heads)
  │── LowRankAdaLN (rank=256) timestep conditioning
  │── JointAttention: self-KV + text-context-KV + speaker-context-KV
  │── Half-RoPE (first half of head_dim rotated, rest passthrough)
  │── SwiGLU MLP (ratio 2.875)
  │
Euler RF ODE Solver (40 steps, CFG)
  │── noise → DAC-VAE latent sequence (32-dim continuous)
  │
Semantic-DACVAE Decoder (48 kHz reconstruction)
  │── 32-dim latent → Snake1d + ConvTranspose1d upsampling → PCM
  │
Output: float32 mono @ 48 kHz

Source model

• Upstream: Aratako/Irodori-TTS-500M-v3 (safetensors, ~1.9 GB)
• Codec: Aratako/Semantic-DACVAE-Japanese-32dim
• Tokenizer: sbintuitions/sarashina2.2-0.5b
• Code: Aratako/Irodori-TTS
• Converted with: models/convert-irodori-tts-to-gguf.py