---
title: Irodori-TTS Compare
emoji: 🗣️
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.6.0
python_version: "3.10"
app_file: app.py
pinned: false
license: mit
short_description: Irodori-TTS v4.1 Small vs Yomi vs Yomi-Tech-tuned
models:
- Aratako/Irodori-TTS-v4.1-Small
- takuma104/Irodori-TTS-v4.1-Small-Yomi
- pentacoxian-dev/Irodori-TTS-v4.1-Small-Yomi-Tech-tuned
- Aratako/Semantic-DACVAE-Japanese-32dim
---

# Irodori-TTS Compare

Hear Irodori-TTS v4.1 Small variants side by side: pick two, type a Japanese sentence (or pick an example),
and both read it with the same seed, voice and settings.

| Variant | What it is |
| :--- | :--- |
| [Irodori-TTS-v4.1-Small](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small) | the upstream model |
| [Irodori-TTS-v4.1-Small-Yomi](https://huggingface.co/takuma104/Irodori-TTS-v4.1-Small-Yomi) | text-encoder fine-tune for kanji readings |
| [Irodori-TTS-v4.1-Small-Yomi-Tech-tuned](https://huggingface.co/pentacoxian-dev/Irodori-TTS-v4.1-Small-Yomi-Tech-tuned) | Yomi plus a technical-term reading frontend stored in the model file (MQTT, API key, schedule, …) |

- **Yomi variant.** It is served from the Yomi-Tech-tuned file with the reading frontend turned off. The
  tensors are identical, and so is the output.
- **Reading table.** Above the audio, a table shows how the frontend will read each English term in the
  text, and whether the reading came from the dictionary, the reader, or letter-by-letter spelling.

## Running

- **On Spaces.** Choose **ZeroGPU** hardware in the Space settings. Both models are downloaded from the
  Hub at startup.
- **Locally.** In a Python 3.10 environment, run `pip install -r requirements.txt gradio==6.6.0`, then
  `python app.py`.
  - To use local files, set `OURS_CHECKPOINT` and `BASE_CHECKPOINT` to `model.safetensors` paths, each
    with its `tokenizer/` folder beside it.
  - `GRADIO_SERVER_PORT` picks the port.

## Code and licences

- `irodori_tts/`: [Aratako/Irodori-TTS](https://github.com/Aratako/Irodori-TTS) at commit `89f9d8f`,
  unmodified. MIT; see `irodori_tts/LICENSE`.
- `tech_reading/` and `app.py`: MIT.
- **Model licences.** The models are downloaded from their own repos and keep their own licences. The
  reading data inside the Yomi-Tech-tuned checkpoint is partly derived from JMdict (CC BY-SA 4.0); see
  that model's card.
- **Ethics.** The [ethical restrictions of Irodori-TTS](https://huggingface.co/Aratako/Irodori-TTS-v4.1-Small#ethical-restrictions)
  apply: no impersonation, and no misleading deepfakes.