---
language:
- ja
license: other
license_link: https://huggingface.co/spaces/litagin/moe-speech-license
task_categories:
- text-to-speech
size_categories:
- 10K<n<100K
tags:
- speech
- audio
- japanese
- tts
- piper
- multi-speaker
- vits
pretty_name: MoeSpeech 20 Speakers (LJSpeech Format)
extra_gated_prompt: |
  You agree to the terms of the [LICENSE](https://huggingface.co/spaces/litagin/moe-speech-license) when using this dataset. Please read it carefully.

  このデータセットを使用する際には、[LICENSE](https://huggingface.co/spaces/litagin/moe-speech-license)の条件に同意したことになります。必ずお読みください。
extra_gated_fields:
  Please provide a detailed description of the specific tasks or applications for which you intend to use this dataset, DO NOT use general statements such as 'for research', 'learning' or 'personal use" (このデータセットを使用する具体的なタスクや用途を、曖昧な表現ではなく（NG例:「研究」「機械学習」「私的利用」）、詳細に記述してください) : text
  I have read and agree to the terms of the LICENSE (利用規約を読んで同意しました): checkbox
---

# MoeSpeech 20 Speakers - LJSpeech Format

[moe-speech-plus-ljspeech](https://huggingface.co/datasets/ayousanz/moe-speech-plus-ljspeech)から発話数上位20話者を抽出したデータセットです。
[Piper TTS](https://github.com/rhasspy/piper)などのTTSモデル学習に最適化されています。

## Dataset Summary

| 項目 | 値 |
|------|-----|
| 話者数 | 20 |
| 総発話数 | 60,233 |
| フォーマット | LJSpeech |
| サンプルレート | 22050 Hz |
| 言語 | 日本語 |
 | 総音声時間     | 約90.6時間 |   

## Speaker Statistics

| Speaker ID | Utterances | Internal ID |
|------------|------------|-------------|
| 940de876 | 4,675 | 0 |
| 2cf01874 | 4,632 | 1 |
| bbd90363 | 3,747 | 2 |
| 1a5a3db8 | 3,295 | 3 |
| 4e2f4ba6 | 3,268 | 4 |
| cc948b89 | 3,043 | 5 |
| ad28b91b | 3,033 | 6 |
| ee093a4f | 2,957 | 7 |
| bce2a5af | 2,926 | 8 |
| 18460462 | 2,861 | 9 |
| 917feebd | 2,811 | 10 |
| b8b5fe66 | 2,769 | 11 |
| 00163dc9 | 2,757 | 12 |
| 773a4156 | 2,749 | 13 |
| 46d6bf83 | 2,654 | 14 |
| 36ea135b | 2,631 | 15 |
| bc778ddb | 2,435 | 16 |
| 6d565f54 | 2,351 | 17 |
| 9febd2ae | 2,327 | 18 |
| 8b6e7173 | 2,312 | 19 |

## Dataset Structure

```
moe-speech-20speakers-ljspeech/
├── metadata.csv      # LJSpeech format: filename|speaker_id|text
├── wavs/            # Audio files (22050Hz mono WAV)
│   ├── {speaker_id}_{utterance_id}.wav
│   └── ...
└── wavs.zip         # Compressed audio files (21GB)
```

### metadata.csv Format

```
{speaker_id}_{utterance_id}|{speaker_id}|{transcription_text}
```

Example:
```
940de876_0000|940de876|こんにちは、今日は良い天気ですね。
```

## Download

### Using huggingface-cli

```bash
# Install huggingface-cli
pip install -U "huggingface_hub[cli]"

# Login
huggingface-cli login

# Download entire dataset
huggingface-cli download --repo-type dataset ayousanz/moe-speech-20speakers-ljspeech --local-dir moe-speech-20speakers-ljspeech
```

### Extracting wavs.zip

```bash
cd moe-speech-20speakers-ljspeech
unzip wavs.zip
```

## Usage with Piper TTS

### Preprocessing

```bash
uv run python -m piper_train.preprocess \
  --language ja \
  --input-dir /path/to/moe-speech-20speakers-ljspeech \
  --output-dir /path/to/dataset-moe-speech-20speakers \
  --dataset-format ljspeech \
  --sample-rate 22050
```

### Training (Multi-GPU)

```bash
NCCL_DEBUG=INFO \
NCCL_P2P_DISABLE=1 \
NCCL_IB_DISABLE=1 \
uv run python -m piper_train \
  --dataset-dir /path/to/dataset-moe-speech-20speakers \
  --accelerator gpu \
  --devices 4 \
  --precision 16-mixed \
  --max_epochs 100 \
  --batch-size 5 \
  --checkpoint-epochs 1 \
  --quality medium \
  --base_lr 2e-4 \
  --ema-decay 0.9995 \
  --num-workers 0 \
  --no-pin-memory \
  --default_root_dir /path/to/output
```

## Related Resources

- **Original Dataset**: [moe-speech-plus-ljspeech](https://huggingface.co/datasets/ayousanz/moe-speech-plus-ljspeech)
- **5 Speakers Dataset**: [moe-speech-5speakers-ljspeech](https://huggingface.co/datasets/ayousanz/moe-speech-5speakers-ljspeech)
- **5 Speakers Trained Model**: [piper-plus-moe-speech-top-5speakers](https://huggingface.co/ayousanz/piper-plus-moe-speech-top-5speakers)
- **Piper TTS**: [https://github.com/rhasspy/piper](https://github.com/rhasspy/piper)

## License

This dataset inherits the license from the original [moe-speech](https://huggingface.co/spaces/litagin/moe-speech-license) dataset.

**Important**: This dataset is only permitted for use under Article 30-4 of the Copyright Law of Japan for data analysis (such as machine learning) purposes. Any use for purposes other than those specified is prohibited by the license.

このデータセットは、著作権法第三十条の四の情報解析（機械学習等）の目的でのみ使用が許可されています。

## Citation

If you use this dataset, please cite the original MoeSpeech dataset:

```
@misc{moespeech,
  author = {litagin},
  title = {MoeSpeech: Japanese Character Voice Dataset},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/litagin/moe-speech}
}
```