Emilia-Dataset
Official dataset for large-scale training of speech models
1 min read
This large-scale multilingual dataset is designed for training and testing text-to-speech and automatic speech recognition systems. It includes over 200,000 hours of audio recordings with corresponding text transcripts across six languages: Chinese, English, Japanese, French, German, and Korean. The dataset consists of an original 101,000-hour base and an additional pool of 114,000 hours of high-quality materials. The resource will be useful for researchers and developers building advanced voice engines and speech assistants.
Key features
- Support for six languages, including English, Chinese, Japanese, and European languages.
- Huge total volume—over 200,000 hours of audio data.
- Open structure and accompanying Emilia-Pipe preprocessing pipeline.
- Distribution of part of the materials under the CC BY-NC 4.0 license.
What you can do
- Train your own state-of-the-art text-to-speech models with high naturalness.
- Fine-tune and evaluate automatic speech recognition systems.
- Use the Emilia-Pipe code for your own preprocessing and filtering of speech audio data.
Example
Downloading the dataset using the datasets library:
from datasets import load_dataset
dataset = load_dataset("amphion/Emilia-Dataset")
print(dataset)Read more
DetailsTotal volume: over 200,000 hours of audio
Languages: zh, en, ja, fr, de, ko
Tools: Emilia-Pipe preprocessing pipeline
Purpose: training TTS and ASR models








Comments