emilia-yodas-de-descriptions

Natural Language Voice Descriptions of the Emilia-YODAS German Dataset

German in-the-wild speech (talk shows, interviews, podcasts, vlogs and other YouTube content) paired with its transcript, acoustic measurements, discrete German descriptor tags and, for female speakers, a free-text German description of the voice and recording conditions. The dataset is intended for training description-conditioned TTS models such as Parler-TTS.

The data was built as part of research work. It is a random subset of the German part of Emilia-YODAS, processed until the female split reached about 350 hours.

Splits

Both splits have the same columns; only the female split is described. The split a clip lands in is decided by a gender classifier (see Columns), not by human labels, so a small share of clips will be in the wrong split.

Sources and attribution

Emilia-YODAS was built by the Emilia authors from YODAS2, which only contains YouTube videos published under a Creative Commons Attribution license. Emilia distributes Emilia-YODAS under CC BY 4.0, and this derived dataset is released under CC BY 4.0 as well. The original Emilia subset (CC BY-NC 4.0) is not included. The copyright of the underlying recordings remains with the original creators; see the Emilia and YODAS dataset cards for their full terms. If you are the owner of a recording included here and would like it removed, please open a discussion on this repository.

The pool was shuffled once with a fixed seed, and rows are taken in that order, so the dataset is a random sample of the German Emilia-YODAS audio. Please cite the source corpora:

@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}
}

@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{li2023yodas,
  title={Yodas: Youtube-Oriented Dataset for Audio and Speech},
  author={Li, Xinjian and Takamichi, Shinnosuke and Saeki, Takaaki and Chen, William and Shiota, Sayaka and Watanabe, Shinji},
  booktitle={2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  pages={1--8},
  year={2023},
  organization={IEEE}
}

Columns

How it was made

The annotation follows the approach of Data-Speech (Lyth & King, 2024, Natural language guidance of high-fidelity text-to-speech with synthetic annotations), adapted for German, with the same pipeline and settings as leonhard-behr/voxpopuli-mls-de-descriptions.

The descriptor bins are reused unchanged from that dataset, so a tag such as "sehr hohe Stimme" or "etwas schnell" means the same measured range in both datasets and they can be combined for training. Because in-the-wild recordings are noisier, more reverberant and faster than audiobooks and parliamentary speech, the tag distributions here are shifted accordingly.

Every 2,000-row shard was checked automatically before upload: rows, transcripts and embedded audio match the source byte for byte, tags recompute exactly from the measurements, and per-stage alignment checks confirm that measurements belong to the right clips.

Limitations

• text_description is LLM-generated and is not always faithful to the tags. For example, a description may state a different speaking rate than the speaking_rate_bin tag. The tag columns are the authoritative annotation.
• gender, and therefore the split, and all quality and acoustic columns are model predictions, not human labels.
• The descriptor bins are relative to the VoxPopuli + MLS German distribution, not to any absolute scale.
• For a small number of very short clips, Brouhaha returns no SNR/C50 (NaN); their noise and reverberation tags fall into the lowest bin and are unreliable.