---
license: cc-by-4.0
task_categories:
- audio-classification
- text-to-speech
- automatic-speech-recognition
language:
- en
- zh
tags:
- speech
- tts
- emilia-yodas
---

**Use this dataset in conjuction with:**

https://github.com/fangningshao/YodasSpeakerPool


## YodasSpeakerPool

YodasSpeakerPool is a curated, richly-annotated multi-speaker dataset featuring 7,600 unique speakers (3.4K Chinese, 4.2K English). 

Derived from the Emilia-YODAS corpus, each sample is annotated by Gemini 2.5 Flash for its vocal characteristics and audio quality.

### Dataset Features

* Audio Specs: 4–15 second WAV samples of clean speech.
* Rich Metadata: Includes ASR transcripts, noise levels, dialects, emotions, style tags, and detailed 200-500 word vocal descriptions for the speakers.
* Diverse Styles: Spans news, podcasts, storytelling, singing, and more.

### Annotation Schema

Each speaker is annotated with the following aspects using Gemini 2.5 Flash:

| Field | Type | Description |
|-------|------|-------------|
| `asr` | String | Transcribed text from the audio (what the speaker is saying) |
| `noise_level` | Integer (0-3) | Audio quality: 0 = clear recording, 1 = slight background noise, 2 = moderate noise, 3 = loud noise |
| `naturalness` | Integer (0-3) | Voice naturalness: 0 = mechanical/robotic, 1 = somewhat natural, 2 = natural, 3 = very natural human voice |
| `language` | String | Language code (e.g., "en-US", "zh-CN", "zh-TW", "ja-JP", "fr-FR") |
| `dialect` | String | Specific dialect or accent (e.g., "American English", "British English", "普通话", "粤语", "上海话", "四川话", "天津话") |
| `style` | String | Speaking style (e.g., "chatting", "news", "podcast", "storytelling", "lecture", "presentation", "reading", "singing") |
| `emotion` | String | Primary emotion (e.g., "neutral", "happy", "sad", "angry", "surprised", "fearful", "excited", "calm", "passionate") |
| `maybe_ai` | Integer (0-1) | AI detection flag: 1 = possibly AI/synthetic voice or reminds of AI assistant, 0 = certainly human |
| `speaker_description` | String | Detailed 200-500 word English description covering voice characteristics, speaker demographics, speaking style, audio quality, suitable use cases, and unique vocal features |

**Note:** The audio understanding quality is determined by the capabilities of the annotation model (Gemini 2.5 Flash). We may update the dataset with improved annotations as models advance in the future.



### Use Cases
- Zero-shot Voice Cloning: High speaker diversity for robust model generalization.
- Multi-speaker TTS: Fine-grained control using style and emotion tags.
- Voice Understanding: Benchmarking audio-to-text descriptive models.


### Citation

If you use **YodasSpeakerPool** in your research, please cite:

```bibtex
@misc{yodasspeakerpool2026,
  title={YodasSpeakerPool: A Richly-Annotated Multi-Speaker Dataset for Voice Cloning},
  author={Shao, Fangning},
  year={2026},
  howpublished={\url{https://github.com/fangningshao/YodasSpeakerPool}},
  note={Built from Emilia-YODAS, annotated with Gemini 2.5 Flash}
}
```


For more reference on how to use the data: https://github.com/fangningshao/YodasSpeakerPool