---
language:
  - ja
task_categories:
  - text-to-speech
tags:
  - liquid-audio
  - tts
  - japanese
  - irodori-tts
license: openrail++
---

# Japanese TTS Dataset for Liquid Audio Fine-tuning

Generated by **Irodori-TTS-600M-v3-VoiceDesign** and formatted for
[LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) fine-tuning.

## Dataset Structure

| Split      | Samples |
|------------|---------|
| train      | 9322 |
| validation | 491 |

## Columns

| Column          | Type   | Description                                      |
|-----------------|--------|--------------------------------------------------|
| `id`            | string | Sample identifier                                |
| `source_type`   | string | `toxic_to_clean` or `identity`                   |
| `system_prompt` | string | Voice style instruction                          |
| `input_text`    | string | Input text (user turn)                           |
| `target_text`   | string | Target text (assistant turn)                     |
| `input_audio`   | binary | Input WAV bytes (Irodori-synthesized)            |
| `target_audio`  | binary | Target WAV bytes ← used in fine-tuning           |
| `topic`         | string | Topic label (identity samples only)              |

> **Note**: audio columns are raw WAV bytes (`Value("binary")`), not `Audio` feature,
> to avoid the `torchcodec` dependency. Decode with `soundfile` in your training code.

## System Prompt

```
Perform TTS in Japanese. Use a natural Japanese female voice.
```

## Usage in step2_train_liquid_audio.py

```python
import io, soundfile as sf

def wav_bytes_to_array(b: bytes):
    arr, sr = sf.read(io.BytesIO(b), dtype="float32", always_2d=False)
    return arr, sr
```