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](https://github.com/sunnyxrxrx/X-Voice).

## 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

| Source | Format | Sample Rate |
|---------------|--------|-------------|
| Emilia | mp3 | 24kHz |
| Gigaspeech2 | flac | 16kHz |
| KoreaSpeech | flac | 16kHz |
| ReazonSpeech | flac | 16kHz |
| LEMAS | mp3 | 16kHz |
| MLS | flac | 16kHz |
| Granary | ogg | 16kHz |

> **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. **Duration & Speaking Rate Filtering**: Remove segments <0.5s or >30s; filter by language-specific speaking rate thresholds.
2. **Language Validation**: Verify text language consistency using `langdetect`.
3. **Deduplication**: Remove duplicate texts appearing more than 20 times to avoid overfitting.
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
```text
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
```
modelscope download --dataset sunnyxrxrx/X-Voice-Dataset-Train --local_dir [local path you want to place the dataset]
```
### Unzip
```bash
cd [local path you want to place the dataset]

for lang in 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](https://github.com/sunnyxrxrx/X-Voice/blob/main/src/x_voice/train/README.md) for subsequent training process.
## 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.

| Dataset | License | Commercial Use |
|---|---|---|
| Multilingual LibriSpeech | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | Yes |
| Emilia | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | Yes |
| LEMAS | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | Yes |
| VoxPopuli | [CC-0](https://creativecommons.org/publicdomain/zero/1.0/) + [European Parliament's legal notice](https://www.europarl.europa.eu/legal-notice/en/) for the raw data | Yes |
| Granary (MOSEL Part) | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | Yes |
| GigaSpeech 2 | [License agreement required](https://huggingface.co/datasets/speechcolab/gigaspeech2) | See terms |
| Reazon Speech | [CDLA-Sharing-1.0 ](https://cdla.dev/sharing-1-0/) + **only for the purpose of [Japanese Copyright Act](https://www.cric.or.jp/english/clj/cl2.html) Article 30-4.** | See terms |
| KoreaSpeech | Refer to the [repo](https://huggingface.co/datasets/jp1924/KoreaSpeech) | |

# X-Voice训练数据集

## 概述
X-Voice训练数据集是一款**大规模多语言语音语料库（large-scale multilingual speech corpus）**，专为高性能语音模型打造，为跨语言语音与韵律建模提供了坚实的基础。同时它也是[X-Voice模型](https://github.com/sunnyxrxrx/X-Voice)的训练集。

## 核心统计数据
- **总语音时长**：42万小时
- **覆盖30种语言**
- **欧洲语言**：bg（保加利亚语）、cs（捷克语）、da（丹麦语）、de（德语）、el（希腊语）、en（英语）、es（西班牙语）、et（爱沙尼亚语）、fi（芬兰语）、fr（法语）、hr（克罗地亚语）、hu（匈牙利语）、it（意大利语）、lt（立陶宛语）、lv（拉脱维亚语）、mt（马耳他语）、nl（荷兰语）、pl（波兰语）、pt（葡萄牙语）、ro（罗马尼亚语）、ru（俄语）、sk（斯洛伐克语）、sl（斯洛文尼亚语）、sv（瑞典语）。
- **亚洲语言**：id（印度尼西亚语）、ja（日语）、ko（韩语）、th（泰语）、vi（越南语）、zh（中文）。

<img src="image-1.png" alt="不同语言的时长统计" width="420" />

## 数据来源
我们聚合了各语言的高质量开源语音数据集：
- **中文与英语**：Emilia
- **越南语、泰语、印度尼西亚语**：GigaSpeech 2
- **韩语**：KoreaSpeech
- **日语**：ReazonSpeech
- **俄语**：LEMAS
- **欧洲语言（西班牙语、意大利语、法语等）**：Multilingual Librispeech（MLS）、Granary

| 数据源 | 格式 | 采样率 |
|---------------|--------|-------------|
| Emilia | mp3 | 24kHz |
| Gigaspeech2 | flac | 16kHz |
| KoreaSpeech | flac | 16kHz |
| ReazonSpeech | flac | 16kHz |
| LEMAS | mp3 | 16kHz |
| MLS | flac | 16kHz |
| Granary | ogg | 16kHz |

> **格式说明**：
> 对于原本以FLAC格式发布的数据集，我们**完整保留其无损FLAC文件，未将其重新压缩为MP3或OGG等有损格式**。

## 处理流程
我们采用严格的多阶段过滤流程以保障数据质量：
1. **时长与语速过滤**：移除时长小于0.5秒或大于30秒的语音片段；依据各语言特定的语速阈值进行过滤。
2. **语言验证**：使用`langdetect`工具验证文本与标注语言的一致性。
3. **去重处理**：移除出现次数超过20次的重复文本，避免模型过拟合。
4. **声学质量管控**：通过DNSMOS语音质量评估模型过滤低质量音频。

## 数据集亮点
- 覆盖多样的语言类型与时间分布
- 经过清洗的高质量语音-文本训练配对样本
- 针对多语言语音建模与泛化能力进行了优化

## 数据组织结构
text
X-Voice-Dataset-Train/
├── tars/ # 语音数据
│ ├── bg/
│ │ ├── bg_vox_part001.tar
│ │ ├── bg_vox_part002.tar
│ │ └── ...
| ├── ...
│ └── zh/
│ ├── zh_emilia_part001.tar
│ ├── zh_emilia_part002.tar
│ └── ...
│
├── csvs/ # 转录数据
│ ├── metadata_bg_voxpopuli.csv
│ ├── ...
│ └── metadata_zh_emilia.csv
│
└── csvs_stage2/ # 第二阶段微调用转录数据
├── metadata_bg_voxpopuli.csv
├── ...
└── metadata_zh_emilia.csv

## 数据集使用
### 命令行（CLI）下载

modelscope download --dataset sunnyxrxrx/X-Voice-Dataset-Train --local_dir [你希望存放数据集的本地路径]

### 数据解压
bash
cd [你希望存放数据集的本地路径]

for lang in 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

随后你可以参考[X-Voice训练指南](https://github.com/sunnyxrxrx/X-Voice/blob/main/src/x_voice/train/README.md)开展后续训练流程。

## 许可证
**本数据集包含多源数据，各子数据集拥有独立的许可证**。用户必须遵守其所使用的各子数据集的许可证条款。

| 数据集 | 许可证 | 商业使用 |
|---|---|---|
| Multilingual LibriSpeech | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | 允许 |
| Emilia | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | 允许 |
| LEMAS | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | 允许 |
| VoxPopuli | [CC-0](https://creativecommons.org/publicdomain/zero/1.0/) + [欧洲议会法律声明](https://www.europarl.europa.eu/legal-notice/en/)（针对原始数据） | 允许 |
| Granary (MOSEL Part) | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) | 允许 |
| GigaSpeech 2 | [需签署许可协议](https://huggingface.co/datasets/speechcolab/gigaspeech2) | 详见协议条款 |
| Reazon Speech | [CDLA-Sharing-1.0](https://cdla.dev/sharing-1-0/) + **仅可用于符合《日本著作权法》第30-4条的使用场景**，相关法律原文可参考[链接](https://www.cric.or.jp/english/clj/cl2.html) | 详见协议条款 |
| KoreaSpeech | 参考[对应仓库](https://huggingface.co/datasets/jp1924/KoreaSpeech) | |