README.md

---
language:
- en
- zh
license: other
license_name: mixed-cc-by-4.0-apache-2.0
task_categories:
- text-to-speech
pretty_name: Multilingual TTS Dataset (LJSpeech Format)
size_categories:
- 10K<n<100K
tags:
- audio
- speech
- tts
- text-to-speech
- multilingual
- english
- chinese
---

# Multilingual TTS Dataset (LJSpeech Format)

A high-quality multilingual Text-to-Speech dataset combining English and Chinese speech data, optimized for TTS training and suitable for commercial use.

## 🎯 Quick Start

```python
from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("ayousanz/multi-dataset-v2")

# Access data
for item in dataset["train"]:
    audio = item["audio"]          # 22050Hz mono audio
    text = item["transcription"]   # Original text
    speaker = item["speaker_id"]   # Speaker identifier  
    language = item["language"]    # "en" or "zh"
```

## 📊 Dataset Statistics

| **Metric** | **Value** |
|------------|-----------|
| **Total Duration** | 97.2 hours |
| **Total Utterances** | 95,568 |
| **Languages** | English, Chinese |
| **Speakers** | 421 unique speakers |
| **Audio Format** | 22050Hz, 16-bit, mono WAV |

### Language Breakdown

| **Language** | **Hours** | **Speakers** | **Utterances** |
|--------------|-----------|--------------|----------------|
| English | 48.6 | 247 | 32,310 |
| Chinese | 48.6 | 174 | 63,258 |

### Duration Distribution

| **Range** | **Count** | **Percentage** |
|-----------|-----------|----------------|
| 0-2s | 28,555 | 29.9% |
| 2-5s | 48,261 | 50.5% |
| 5-10s | 14,167 | 14.8% |
| 10-15s | 3,417 | 3.6% |
| 15-20s | 1,168 | 1.2% |
| 20s+ | 0 | 0.0% |

## 📁 Repository Structure

```
├── audio/                          # Audio files (ZIP compressed)
│   ├── train_english.zip          # English training audio
│   ├── train_chinese.zip          # Chinese training audio  
│   ├── validation_english.zip     # English validation audio
│   ├── validation_chinese.zip     # Chinese validation audio
│   ├── test_english.zip           # English test audio
│   └── test_chinese.zip           # Chinese test audio
├── metadata/                       # Metadata files
│   ├── train.csv                  # Training metadata (all languages)
│   ├── validation.csv             # Validation metadata (all languages)
│   └── test.csv                   # Test metadata (all languages)
├── dataset_info.json              # Dataset statistics and info
├── multilingual_tts_ljspeech.py   # Dataset loader script
└── README.md                      # This file
```

## 💾 Download Instructions

### Option 1: Using Hugging Face CLI (Recommended)

```bash
# Install Hugging Face CLI
pip install huggingface-hub

# Download entire dataset
huggingface-cli download ayousanz/multi-dataset-v2 --repo-type dataset --local-dir ./multilingual-tts

# Download specific files only
huggingface-cli download ayousanz/multi-dataset-v2 audio/train_english.zip metadata/train.csv --repo-type dataset --local-dir ./multilingual-tts
```

### Option 2: Using Python

```python
from huggingface_hub import snapshot_download

# Download entire dataset
snapshot_download(
    repo_id="ayousanz/multi-dataset-v2",
    repo_type="dataset", 
    local_dir="./multilingual-tts"
)
```

### Extracting Audio Files

After downloading, extract the ZIP files:

```bash
cd multilingual-tts
for zip_file in audio/*.zip; do
    unzip "$zip_file" -d audio_extracted/
done
```

## 🚀 Usage Examples

### Basic Usage

```python
from datasets import load_dataset

dataset = load_dataset("ayousanz/multi-dataset-v2")

# Filter by language
english_data = dataset["train"].filter(lambda x: x["language"] == "en")
chinese_data = dataset["train"].filter(lambda x: x["language"] == "zh")

# Filter by speaker
speaker_data = dataset["train"].filter(lambda x: x["speaker_id"] == "en_1234")
```

### For TTS Training

```python
# Example with PyTorch DataLoader
from torch.utils.data import DataLoader

def collate_fn(batch):
    audios = [item["audio"]["array"] for item in batch]
    texts = [item["transcription"] for item in batch]
    speakers = [item["speaker_id"] for item in batch]
    return audios, texts, speakers

dataloader = DataLoader(
    dataset["train"], 
    batch_size=32, 
    collate_fn=collate_fn,
    shuffle=True
)
```

## 📋 Data Format

Each sample contains:

- **`audio_id`**: Unique identifier for the audio file
- **`audio`**: Audio data (22050Hz, 16-bit, mono)
- **`transcription`**: Original text transcription
- **`normalized_text`**: Normalized text for TTS training
- **`speaker_id`**: Speaker identifier with language prefix (`en_*` or `zh_*`)
- **`language`**: Language code (`en` for English, `zh` for Chinese)

## 📜 License

This dataset combines data from multiple sources:

- **English data (LibriTTS-R)**: CC BY 4.0 - requires attribution
- **Chinese data (AISHELL-3)**: Apache 2.0

### Attribution Requirements

When using this dataset, please cite:

```bibtex
@dataset{{multilingual_tts_ljspeech,
  title={{Multilingual TTS Dataset in LJSpeech Format}},
  year={{2024}},
  note={{English: LibriTTS-R (CC BY 4.0), Chinese: AISHELL-3 (Apache 2.0)}}
}}
```

## 🔗 Source Datasets

- **LibriTTS-R**: https://openslr.org/141/
- **AISHELL-3**: https://openslr.org/93/

## ⚡ Performance Notes

- Audio files are stored in ZIP format for faster download
- Use `datasets` library's built-in caching for optimal performance
- Consider using `streaming=True` for large-scale training to save memory

## 🤝 Contributing

Found an issue? Please report it on the [repository issues page](https://huggingface.co/datasets/ayousanz/multi-dataset-v2/discussions).