README.md

---
license: cc-by-sa-4.0
language:
  - ja
  - en
  - zh
  - es
  - fr
  - pt
task_categories:
  - text-to-speech
tags:
  - tts
  - vits
  - multilingual
  - piper
  - webdataset
size_categories:
  - 100K<n<1M
---

# Piper Plus Base Dataset (6-Language Multilingual)

[Piper Plus](https://github.com/ayutaz/piper-plus) TTS の事前学習用 6 言語マルチリンガルデータセットです。音素化・スペクトログラム計算済みのため、ダウンロード後すぐに学習を開始できます。

## Dataset Summary

| Item | Value |
|------|-------|
| Total utterances | 497,519 |
| Speakers | 571 |
| Languages | 6 |
| Phoneme symbols | 173 |
| Sample rate | 22,050 Hz |
| Format | WebDataset (tar shards) |
| Shards | 48 |
| Total size | ~411 GB |

### Per-Language Breakdown

| Language | Code | ID | Speakers | Utterances | Source |
|----------|------|----|----------|------------|--------|
| Japanese | ja | 0 | 20 | 59,694 | [MOE-Speech](https://huggingface.co/datasets/ayousanz/moe-speech-20speakers-ljspeech) |
| English | en | 1 | 310 | 64,698 | [LibriTTS-R](https://www.openslr.org/141/) |
| Chinese | zh | 2 | 142 | 63,223 | [AISHELL-3](https://www.openslr.org/93/) (Apache-2.0) |
| Spanish | es | 3 | 63 | 168,374 | [CML-TTS](https://www.openslr.org/146/) (CC-BY-4.0) |
| French | fr | 4 | 28 | 107,464 | [CML-TTS](https://www.openslr.org/146/) (CC-BY-4.0) |
| Portuguese | pt | 5 | 8 | 34,066 | [CML-TTS](https://www.openslr.org/146/) (CC-BY-4.0) |

## Data Format

WebDataset tar 形式。各 tar シャードには以下の 3 ファイルがサンプルごとに含まれます:

```
000000000.audio_norm.pt    # 正規化音声波形 (torch.float32, shape: [1, T])
000000000.audio_spec.pt    # メルスペクトログラム (torch.float16, shape: [513, M])
000000000.json             # メタデータ (下記参照)
```

### Metadata Schema (JSON)

```json
{
  "text": "こんにちは",
  "speaker": "00163dc9",
  "speaker_id": 12,
  "language_id": 0,
  "phonemes": ["k", "o", "N_n", ...],
  "phoneme_ids": [1, 0, 32, 0, 15, ...],
  "prosody_features": [null, null, {"a1": 0, "a2": 1, "a3": 3}, ...]
}
```

| Field | Type | Description |
|-------|------|-------------|
| `text` | string | 元テキスト |
| `speaker` | string | 話者名/ID |
| `speaker_id` | int | 話者インデックス (0-570) |
| `language_id` | int | 言語インデックス (0-5) |
| `phonemes` | list[str] | IPA 音素トークン列 |
| `phoneme_ids` | list[int] | モデル入力用の整数 ID 列 (BOS/EOS/padding 含む) |
| `prosody_features` | list | A1/A2/A3 韻律特徴 (日本語のみ、他言語は null) |

## Audio Specifications

| Parameter | Value |
|-----------|-------|
| Sample rate | 22,050 Hz |
| FFT size | 1,024 |
| Hop length | 256 |
| Window length | 1,024 |
| Mel channels | 513 |
| Spectrogram dtype | float16 |
| Audio dtype | float32 |

## Usage

### Download

```bash
# huggingface-cli でダウンロード
huggingface-cli download ayousanz/piper-plus-base-dataset \
  --repo-type dataset \
  --local-dir ./piper-plus-base-dataset
```

### Load with WebDataset

```python
import webdataset as wds
import torch
import json

dataset = (
    wds.WebDataset("piper-plus-base-dataset/data/train-{00000..00047}-of-00048.tar")
    .decode()
    .map(lambda sample: {
        "audio_norm": torch.load(sample["audio_norm.pt"], weights_only=True),
        "audio_spec": torch.load(sample["audio_spec.pt"], weights_only=True),
        "meta": json.loads(sample["json"]),
    })
)

for sample in dataset:
    print(sample["meta"]["text"], sample["audio_norm"].shape)
    break
```

### Extract to Piper Training Format

tar シャードを展開し、Piper の学習形式 (`dataset.jsonl` + `cache/`) に変換するには:

```python
import io
import json
import tarfile
from pathlib import Path

output_dir = Path("./extracted_dataset")
cache_dir = output_dir / "cache" / "22050"
cache_dir.mkdir(parents=True, exist_ok=True)

entries = []
for shard in sorted(Path("./piper-plus-base-dataset/data").glob("*.tar")):
    with tarfile.open(shard) as tf:
        members = sorted(tf.getnames())
        keys = sorted(set(m.split(".")[0] for m in members))
        for key in keys:
            meta = json.loads(tf.extractfile(f"{key}.json").read())
            # Save .pt files to cache
            for ext in ["audio_norm.pt", "audio_spec.pt"]:
                data = tf.extractfile(f"{key}.{ext}").read()
                (cache_dir / f"{key}.{ext.replace('audio_norm', 'norm').replace('audio_spec', 'spec')}").write_bytes(data)
            meta["audio_norm_path"] = str(cache_dir / f"{key}.norm.pt")
            meta["audio_spec_path"] = str(cache_dir / f"{key}.spec.pt")
            entries.append(meta)

with open(output_dir / "dataset.jsonl", "w") as f:
    for e in entries:
        f.write(json.dumps(e, ensure_ascii=False) + "\n")
```

## Training

このデータセットで Piper Plus を学習する例:

```bash
python -m piper_train \
  --dataset-dir ./extracted_dataset \
  --prosody-dim 16 \
  --accelerator gpu --devices 4 --precision 32-true \
  --max_epochs 75 --batch-size 20 --samples-per-speaker 2 \
  --checkpoint-epochs 5 --quality medium \
  --base_lr 2e-4 --disable_auto_lr_scaling \
  --ema-decay 0.9995 \
  --max-phoneme-ids 400 \
  --no-wavlm \
  --audio-log-epochs 5
```

## Pre-trained Models

| Model | Description | Link |
|-------|-------------|------|
| Base (6-lang, 75 epoch) | 571 speakers, HiFi-GAN decoder | [ayousanz/piper-plus-base](https://huggingface.co/ayousanz/piper-plus-base) |
| Tsukuyomi-chan | Single speaker finetuned from base | [ayousanz/piper-plus-tsukuyomi-chan](https://huggingface.co/ayousanz/piper-plus-tsukuyomi-chan) |

## Preprocessing Pipeline

このデータセットは以下のパイプラインで作成されました:

1. **prepare_multilingual_dataset.py** -- 6 言語ソースデータの統合・音素化・音声キャッシュ生成
2. **Speaker filtering** -- 30 発話未満の話者を除外 (657 -> 571 speakers)
3. **Speaker ID remapping** -- 連続 ID (0-570) にリマッピング

## License

- Dataset: CC-BY-SA-4.0
- Source corpora:
  - MOE-Speech: CC-BY-SA-4.0
  - LibriTTS-R: CC-BY-4.0
  - AISHELL-3: Apache-2.0
  - CML-TTS (ES/FR/PT): CC-BY-4.0

## Citation

```bibtex
@misc{piper-plus-base-dataset,
  title={Piper Plus Base Dataset: 6-Language Multilingual TTS Training Data},
  author={ayousanz},
  year={2026},
  url={https://huggingface.co/datasets/ayousanz/piper-plus-base-dataset}
}
```