---
title: Irodori-TTS-v4-Small-fp16
canonical_url: "https://www.modelscope.cn/models/mlx-community/Irodori-TTS-v4-Small-fp16"
md_url: "https://www.modelscope.cn/models/mlx-community/Irodori-TTS-v4-Small-fp16.md"
repository: mlx-community/Irodori-TTS-v4-Small-fp16
last_updated: 2026-08-10
pipeline_tag: text-to-speech
tasks:
  - text-to-speech
model_type:
  - irodori_tts
base_model:
  - Aratako/Irodori-TTS-v4-Small
base_model_relation: finetune
parameters: 766.1M
tensor_type:
  - F32
  - F16
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 6
stars: 0
tags:
  - mlx
  - japanese
---

# Irodori-TTS-v4-Small-fp16

> Irodori-TTS-v4-Small-fp16 - mlx-community 在 ModelScope 开源的模型。mlx-community/Irodori-TTS-v4-Small-fp16

mlx-community/Irodori-TTS-v4-Small-fp16 是 ModelScope 魔搭社区上的 766.1M 参数text-to-speech模型，基于 Aratako/Irodori-TTS-v4-Small 构建。

- **Repository**: mlx-community/Irodori-TTS-v4-Small-fp16
- **Tasks**: text-to-speech
- **Parameters**: 766.1M
- **Base model**: Aratako/Irodori-TTS-v4-Small
- **Tags**: mlx, japanese
- **Downloads**: 6
- **Stars**: 0
- **Last updated**: 2026-08-10

Source: https://www.modelscope.cn/models/mlx-community/Irodori-TTS-v4-Small-fp16

---

# mlx-community/Irodori-TTS-v4-Small-fp16

This model was converted to MLX format from [Aratako/Irodori-TTS-v4-Small](https://huggingface.co/Aratako/Irodori-TTS-v4-Small) using [mlx-audio version **0.4.7**](https://github.com/Blaizzy/mlx-audio).

## Use with mlx-audio

```bash
pip install -U mlx-audio
```

### Command line

```bash
mlx_audio.tts.generate --model mlx-community/Irodori-TTS-v4-Small-fp16 --text "こんにちは、Irodori TTSのMLX版です。" --ref_audio reference.wav
```

### Python

```python
from mlx_audio.tts.generate import generate_audio

generate_audio(
    text="こんにちは、Irodori TTSのMLX版です。",
    model="mlx-community/Irodori-TTS-v4-Small-fp16",
    ref_audio="reference.wav",
    file_prefix="output",
)
```

## Notes

- Unified v4 model: voice cloning, VoiceDesign (caption) and automatic duration
  prediction in one checkpoint.
- The ModernBERT-ja-310m text encoder weights and its tokenizer are bundled, and
  the Semantic-DACVAE-Japanese-32dim codec ships in `dacvae/`, so inference needs
  no additional downloads.
- Reference audio: up to 120s. Passing a list of clips encodes each separately
  and concatenates them, which matches training better than one long recording.
- With a caption but **no** reference audio, the duration predictor roughly
  doubles the length of short texts (under ~7 tokens) and the model reads the
  sentence twice to fill the window. This matches the reference PyTorch
  implementation. Pass reference audio, or set `duration_scale` / `seconds`.