---
license: mit
language:
- ja
tags:
- text-to-speech
- gguf
---

# Irodori-TTS v4 Small GGUF

Follow [audio.cpp](https://github.com/0xShug0/audio.cpp) for runtime support and updates.

This directory contains the RF, Anime, and MeanFlow variants.
The MeanFlow files are converted from [Aratako/Irodori-TTS-v4.1-Small-MF](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small-MF/tree/ccc78f5d480b6e51b69b2d5042a14c4da04fea6e).
The model, tokenizer, configuration, and DACVAE codec are contained in each GGUF.

| File | Precision | Sampler |
|---|---|---|
| `irodori-tts-v4-small-q8_0.gguf` | Q8_0 | RF |
| `irodori-tts-v4-small-f16.gguf` | F16 | RF |
| `irodori-tts-v4.1-anime-q8_0.gguf` | Q8_0 | RF |
| `irodori-tts-v4.1-small-mf-q8_0.gguf` | Q8_0 | MeanFlow |
| `irodori-tts-v4.1-small-mf-f16.gguf` | F16 | MeanFlow |

The MF files require audio.cpp with Irodori MeanFlow support. The checkpoint selects the
MeanFlow sampler automatically, using four steps by default. RF checkpoints
continue to use their own sampler.

```bash
audiocpp_cli --task tts --family irodori_tts \
  --model irodori-tts-v4.1-small-mf-q8_0.gguf \
  --backend cuda --text "こんにちは、私はAIです。これは音声合成のテストです。" \
  --request-option no_ref=true --out speech.wav
```

For cloning, use `--task clon --voice-ref reference.wav` without `no_ref=true`.
Voice design uses `--request-option instruct="..."` with `no_ref=true`.
`num_inference_steps` overrides the step count. Runtime CFG and Sway controls
do not apply to this distilled checkpoint.

See the [audio.cpp model guide](https://github.com/0xShug0/audio.cpp/blob/main/docs/models/irodori_tts.md)
and [upstream model card](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small-MF)
for usage and model limitations.