ayousanz/Emilia-Dataset-JA-Plus

Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation

This is the official repository 👑 for the Emilia dataset and the source code for the Emilia-Pipe speech data preprocessing pipeline.

News 🔥

• 2024/08/28: Welcome to join Amphion's Discord channel to stay connected and engage with our community!
• 2024/08/27: The Emilia dataset is now publicly available! Discover the most extensive and diverse speech generation dataset with 101k hours of in-the-wild speech data now at HuggingFace or OpenDataLab! 👑👑👑
• 2024/07/08: Our preprint paper is now available! 🔥🔥🔥
• 2024/07/03: We welcome everyone to check our homepage for our brief introduction for Emilia dataset and our demos!
• 2024/07/01: We release of Emilia and Emilia-Pipe! We welcome everyone to explore it on our GitHub! 🎉🎉🎉

Emilia Overview ⭐️

The Emilia dataset is a comprehensive, multilingual dataset with the following features:

• containing over 101k hours of speech data;
• covering six different languages: English (En), Chinese (Zh), German (De), French (Fr), Japanese (Ja), and Korean (Ko);
• containing diverse speech data with various speaking styles from diverse video platforms and podcasts on the Internet, covering various content genres such as talk shows, interviews, debates, sports commentary, and audiobooks.

The table below provides the duration statistics for each language in the dataset.

Language	Duration (hours)
English	46,828
Chinese	49,922
German	1,590
French	1,381
Japanese	1,715
Korean	217

The Emilia-Pipe is the first open-source preprocessing pipeline designed to transform raw, in-the-wild speech data into high-quality training data with annotations for speech generation. This pipeline can process one hour of raw audio into model-ready data in just a few minutes, requiring only the raw speech data.

Detailed descriptions for the Emilia and Emilia-Pipe can be found in our paper.

Emilia Dataset Usage 📖

Emilia is publicly available at HuggingFace.

If you are from mainland China or having a connecting issue with HuggingFace, you can also download Emilia from OpenDataLab.

• To download from HuggingFace:

  1. Gain access to the dataset and get the HF access token from: https://huggingface.co/settings/tokens.
  2. Install dependencies and login HF:
     • Install Python
     • Run pip install librosa soundfile datasets huggingface_hub[cli]
     • Login by huggingface-cli login and paste the HF access token. Check here for details.
  3. Use following code to load Emilia:

  from datasets import load_dataset
  dataset = load_dataset("amphion/Emilia-Dataset", streaming=True)
  print(dataset)
  print(next(iter(dataset['train'])))
  

• To download from OpenDataLab (i.e., OpenXLab), please follow the guidance here to gain access.

ENJOY USING EMILIA!!! 🔥

Use cases

If you want to load a subset of Emilia, e.g., only language DE, you can use the following code:

from datasets import load_dataset
path = "DE/*.tar"
dataset = load_dataset("amphion/Emilia-Dataset", data_files={"de": path}, split="de", streaming=True)
print(dataset) # here should only shows 90 n_shards instead of 2360
print(next(iter(dataset['train'])))

If you want to download all files to your local before using Emilia, remove the streaming=True argument:

from datasets import load_dataset
dataset = load_dataset("amphion/Emilia-Dataset") # prepare 2.4TB space to store Emilia
print(dataset)

Re-build or Processing your own data

If you wish to re-build Emilia from scratch, you may download the raw audio files from the provided URL list and use our open-source Emilia-Pipe preprocessing pipeline to preprocess the raw data. Additionally, users can easily use Emilia-Pipe to preprocess their own raw speech data for custom needs. By open-sourcing the Emilia-Pipe code, we aim to enable the speech community to collaborate on large-scale speech generation research.

Notes

Please note that Emilia does not own the copyright to the audio files; the copyright remains with the original owners of the videos or audio. Users are permitted to use this dataset only for non-commercial purposes under the CC BY-NC-4.0 license.

Emilia Dataset Structure ⛪️

Structure on HuggingFace

On HuggingFace, Emilia is now formatted as WebDataset.

Each audio is tared with a corresponding JSON file (having the same prefix filename) within 2360 tar files.

By utilizing WebDataset, you can easily stream audio data, which is magnitude faster than reading separate data files one by one.

Read the Emilia Dataset Usage 📖 part for a detailed usage guide.

Learn more about WebDataset here.

PS: If you want to download the OpenDataLab format from HuggingFace, you can specify the revision argument to fc71e07e8572f5f3be1dbd02ed3172a4d298f152, which is the old format.

Structure on OpenDataLab

On OpenDataLab, Emilia is formatted using the following structure.

Structure example:

|-- openemilia_all.tar.gz (all .JSONL files are gzipped with directory structure in this file)
|-- EN (114 batches)
|   |-- EN_B00000.jsonl
|   |-- EN_B00000 (= EN_B00000.tar.gz)
|   |   |-- EN_B00000_S00000
|   |   |   `-- mp3
|   |   |       |-- EN_B00000_S00000_W000000.mp3
|   |   |       `-- EN_B00000_S00000_W000001.mp3
|   |   |-- ...
|   |-- ...
|   |-- EN_B00113.jsonl
|   `-- EN_B00113
|-- ZH (92 batches)
|-- DE (9 batches)
|-- FR (10 batches)
|-- JA (7 batches)
|-- KO (4 batches)

JSONL files example:

{"id": "EN_B00000_S00000_W000000", "wav": "EN_B00000/EN_B00000_S00000/mp3/EN_B00000_S00000_W000000.mp3", "text": " You can help my mother and you- No. You didn't leave a bad situation back home to get caught up in another one here. What happened to you, Los Angeles?", "duration": 6.264, "speaker": "EN_B00000_S00000", "language": "en", "dnsmos": 3.2927}
{"id": "EN_B00000_S00000_W000001", "wav": "EN_B00000/EN_B00000_S00000/mp3/EN_B00000_S00000_W000001.mp3", "text": " Honda's gone, 20 squads done. X is gonna split us up and put us on different squads. The team's come and go, but 20 squad, can't believe it's ending.", "duration": 8.031, "speaker": "EN_B00000_S00000", "language": "en", "dnsmos": 3.0442}

Reference 📖

If you use the Emilia dataset or the Emilia-Pipe pipeline, please cite the following papers:

@inproceedings{emilia,
    author={He, Haorui and Shang, Zengqiang and Wang, Chaoren and Li, Xuyuan and Gu, Yicheng and Hua, Hua and Liu, Liwei and Yang, Chen and Li, Jiaqi and Shi, Peiyang and Wang, Yuancheng and Chen, Kai and Zhang, Pengyuan and Wu, Zhizheng},
    title={Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation},
    booktitle={Proc.~of SLT},
    year={2024}
}

@inproceedings{amphion,
    author={Zhang, Xueyao and Xue, Liumeng and Gu, Yicheng and Wang, Yuancheng and Li, Jiaqi and He, Haorui and Wang, Chaoren and Song, Ting and Chen, Xi and Fang, Zihao and Chen, Haopeng and Zhang, Junan and Tang, Tze Ying and Zou, Lexiao and Wang, Mingxuan and Han, Jun and Chen, Kai and Li, Haizhou and Wu, Zhizheng},
    title={Amphion: An Open-Source Audio, Music and Speech Generation Toolkit},
    booktitle={Proc.~of SLT},
    year={2024}
}