XRXRX/X-Voice-Dataset-Train

X-Voice Training Dataset

Overview

The X-Voice training dataset is a large-scale multilingual speech corpus curated for high-performance speech models. It provides a robust foundation for cross-lingual phonetic and prosodic modeling.

Also the train set of X-Voice Model.

Core Statistics

• —Total Speech Duration: 420K hours
• —30 languages
• —European: bg (Bulgarian), cs (Czech), da (Danish), de (German), el (Greek), en (English), es (Spanish), et (Estonian), fi (Finnish), fr (French), hr (Croatian), hu (Hungarian), it (Italian), lt (Lithuanian), lv (Latvian), mt (Maltese), nl (Dutch), pl (Polish), pt (Portuguese), ro (Romanian), ru (Russian), sk (Slovak), sl (Slovenian), sv (Swedish).
• —Asian: id (Indonesian), ja (Japanese), ko (Korean), th (Thai), vi (Vietnamese), zh (Chinese).

<img src="image-1.png" alt="Duration Statistics of Different Languages" width="420" />

Data Sources

We aggregate high-quality open-source speech datasets across languages:

• —Chinese & English: Emilia
• —Vietnamese, Thai, Indonesian: GigaSpeech 2
• —Korean: KoreaSpeech
• —Japanese: ReazonSpeech
• —Russian: LEMAS
• —European Languages (Spanish, Italian, French, etc.): Multilingual Librispeech (MLS), Granary

Format Note: For datasets originally distributed in FLAC format, we retain the lossless FLAC files in their original state, without recompressing them into lossy formats like MP3 or OGG.

Processing Pipeline

A rigorous multi-stage filtering pipeline is applied to ensure data quality:

1. 1.Duration & Speaking Rate Filtering: Remove segments <0.5s or >30s; filter by language-specific speaking rate thresholds.
2. 2.Language Validation: Verify text language consistency using langdetect.
3. 3.Deduplication: Remove duplicate texts appearing more than 20 times to avoid overfitting.
4. 4.Acoustic Quality Control: Filter low-quality audio via DNSMOS speech quality assessment.

Highlights

• —Diverse linguistic and temporal distribution
• —High-quality cleaned speech-text training pairs
• —Optimized for multilingual speech modeling and generalization

Data Structure

X-Voice-Dataset-Train/
├── tars/                     # Speech Data
│   ├── bg/
│   │   ├── bg_vox_part001.tar
│   │   ├── bg_vox_part002.tar
│   │   └── ...
|   ├── ...
│   └── zh/
│       ├── zh_emilia_part001.tar
│       ├── zh_emilia_part002.tar
│       └── ...
│
├── csvs/                     # Tramscript Data
│   ├── metadata_bg_voxpopuli.csv
│   ├── ...
│   └── metadata_zh_emilia.csv
│
└── csvs_stage2/              # Transcript Data for Stage 2 Finetuning
    ├── metadata_bg_voxpopuli.csv
    ├── ...
    └── metadata_zh_emilia.csv

Use the Dataset

CLI Download

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="XRXRX/X-Voice-Dataset-Train",
    # allow_patterns=["data/tars/bg/*.tar"], # you can download a specific package.  
    local_dir=[local path you want to place the dataset],
    repo_type="dataset",
    local_dir_use_symlinks=False
)

Unzip

cd [local path you want to place the dataset]

for lang in data/tars/*/; do
    lang_name=$(basename "$lang")
    mkdir -p "wavs/$lang_name"
    tar xf "$lang"*.tar -C "wavs/$lang_name" --strip-components=1 --skip-old-files
done

Then you can refer to X-Voice Training for subsequent training process.

Note: The data from Gigaspeech 2 (including Thai, Indonesian, and Vietnamese) is currently not included in this repo due to storage constraints and license restrictions. You can download them from the official source or refer to our repo in ModelScope.

License

This dataset contains data from multiple sources, each with its own license. Users must comply with the license of each individual sub-dataset they use.