meetween/mumospee_emilia

Dataset Summary

This dataset is a modified version of the Emilia corpus, converted into parquet format to facilitate optimized I/O operations in high-performance and distributed computing environments. 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.

This release covers the core Emilia corpus (the six languages above); the separate Emilia-YODAS collection is not included. Both the audio shards and the segment metadata are kept aligned, so every metadata row points to a shard present in this repository.

---

Source Data

• Original Dataset: Emilia
• License: This derived dataset is shared under the same license (CC-BY-NC-4.0), with modifications only to format for efficiency.

Modifications

• Data Format: Converted from the original WebDataset .tar archives to parquet format to enhance I/O performance for distributed training, reducing latency during data loading and retrieval.
• Efficiency Optimization: Restructured for reduced storage footprint and faster I/O on high-performance clusters by leveraging parquet’s efficient compression and columnar storage.
• Scope: Restricted to the six core Emilia languages (de, en, fr, ja, ko, zh), so the audio and metadata folders stay aligned.

Dataset Structure

• File Format: Parquet files.
• Languages: 6 languages (de, en, fr, ja, ko, zh).
• Audio Sampling Rate: 24 kHz, MP3 encoded.

The repository contains two top-level folders:

emilia-parquet/ — segment audio + per-segment fields

Sharded per language at emilia-parquet/{lang}/{LANG}-B{nnnnnn}.tar.parquet (one converted parquet per original Emilia .tar). Each row is one utterance-level segment with its MP3 audio embedded; a language's segments are striped across that language's shards.

column	type	description
__key__	string	sample id, e.g. EN_B00000_S00000_W000000
__url__	string	provenance: source tar in amphion/Emilia-Dataset
mp3	struct<array: binary, path: string, sampling_rate: int64>	array = raw MP3 bytes, path = <key>.mp3, sampling_rate = 24000
dnsmos	double	DNSMOS P.835 OVRL audio-quality score
duration	double	segment duration in seconds
id	string	sample id (same as __key__)
language	string	language code
speaker	string	speaker / session id, e.g. EN_B00000_S00000
text	string	transcript
wav	string	relative segment path, e.g. EN_B00000/EN_B00000_S00000/mp3/EN_B00000_S00000_W000000.mp3

metadata/ — per-segment metadata

One file per language at metadata/emilia_<LANG>.parquet. Each row is one Emilia segment.

column	type	description
path	string	segment path (matches the wav field of the audio shard)
url	string	URL of the emilia-parquet/ shard that holds this segment's audio
type	string	constant "audio"
duration	float64	segment duration in seconds
language	string	language code
transcript	string	transcript
tag	string	constant "Emilia"
split	string	constant "train"
license	string	constant "CC-BY-NC-4.0"

Totals: 40,237,814 segments across 6 languages, 101,585 hours of audio, in 2,360 audio shards (2.45 TB).

Statistics

Overall

• 6 languages, 40,237,814 segments, ~101,585 hours of audio.
• 2,360 audio shards (~2.45 TB), MP3 at 24 kHz.
• Average segment duration: 9.09 s.
• Chinese (zh) and English (en) dominate (~49 % and ~46 % of hours); de, fr, ja, ko together make up under 5 %.

Per language

lang	rows	hours	shards	audio size	avg segment (s)	% hours
de	653,109	1,590.5	90	35.1 GB	8.77	1.57 %
en	18,109,854	46,757.7	1,140	1085.2 GB	9.29	46.03 %
fr	543,710	1,381.7	100	60.5 GB	9.15	1.36 %
ja	869,662	1,715.5	70	81.0 GB	7.10	1.69 %
ko	92,182	217.2	40	4.7 GB	8.48	0.21 %
zh	19,969,297	49,922.5	920	1186.1 GB	9.00	49.14 %
total	40,237,814	101,585.1	2,360	~2.45 TB	9.09	100 %

Usage

This dataset is ideal for use in large-scale multilingual speech-to-text and text-to-speech tasks, especially in distributed and high-performance computing environments. The parquet format enhances usability by minimizing I/O overhead, making it well-suited for high-throughput training.

Attribution

This dataset is based on the original Emilia dataset, with modifications for I/O optimization by converting to parquet format. Please cite the original Emilia dataset in any publications or projects using this dataset.

Changelog

Version 1.1 (2026-06-25)

• Recovered the 62 missing Japanese shards (Ja was previously 8/70); ja/ is now complete at 70 shards, rebuilt from amphion/Emilia-Dataset.
• Rebuilt 4 corrupt English shards (EN-B000160, EN-B000370, EN-B000501, EN-B001100) that had unreadable parquet footers.
• Renamed the English audio folder english/ → en/ so it matches the other languages and the metadata URLs.
• Replaced the per-language CSV metadata with regenerated parquet metadata, sourced directly from the audio shards: corrected per-row download urls, clean UTF-8 transcripts, duration stored as float64, and a fixed license value of CC-BY-NC-4.0.

Citation

@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{emilialarge,
    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: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation},
    booktitle={arXiv:2501.15907},
    year={2025}
}