# NonverbalTTS Dataset 🎵🗣️

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.15274617.svg)](https://doi.org/10.5281/zenodo.15274617)
[![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-blue)](https://huggingface.co/datasets/deepvk/NonverbalTTS)

**NonverbalTTS** is a 17-hour open-access English speech corpus with aligned text annotations for **nonverbal vocalizations (NVs)** and **emotional categories**, designed to advance expressive text-to-speech (TTS) research.

## Key Features ✨

- **17 hours** of high-quality speech data
- **10 NV types**: Breathing, laughter, sighing, sneezing, coughing, throat clearing, groaning, grunting, snoring, sniffing
- **8 emotion categories**: Angry, disgusted, fearful, happy, neutral, sad, surprised, other
- **Diverse speakers**: 2296 speakers (60% male, 40% female)
- **Multi-source**: Derived from [VoxCeleb](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/) and [Expresso](https://speechbot.github.io/expresso/) corpora
- **Rich metadata**: Emotion labels, NV annotations, speaker IDs, audio quality metrics

<!-- ## Dataset Structure 📂

NonverbalTTS/
├── wavs/ # Audio files (16-48kHz WAV format)
│ ├── ex01_sad_00265.wav
│ └── ...
├── .gitattributes
├── README.md
└── metadata.csv # Metadata annotations -->


## Metadata Schema (`metadata.csv`) 📋

| Column | Description | Example |
|--------|-------------|---------|
| `index` | Unique sample ID | `ex01_sad_00265` |
| `file_name` | Audio file path | `wavs/ex01_sad_00265.wav` |
| `Emotion` | Emotion label | `sad` |
| `Initial text` | Raw transcription | `"So, Mom, 🌬️ how've you been?"` |
| `Annotator response {1,2,3}` | Refined transcriptions | `"So, Mom, how've you been?"` |
| `Result` | Final fused transcription | `"So, Mom, 🌬️ how've you been?"` |
| `dnsmos` | Audio quality score (1-5) | `3.936982` |
| `duration` | Audio length (seconds) | `3.6338125` |
| `speaker_id` | Speaker identifier | `ex01` |
| `data_name` | Source corpus | `Expresso` |
| `gender` | Speaker gender | `m` |

**NV Symbols**: 🌬️=Breath, 😂=Laughter, etc. (See [Annotation Guidelines](https://zenodo.org/records/15274617))

## Loading the Dataset 💻

```python
from datasets import load_dataset

dataset = load_dataset("deepvk/NonverbalTTS")
```

<!-- # Access train split
```print(dataset["train"][0])```

# Output: {'index': 'ex01_sad_00265', 'file_name': 'wavs/ex01_sad_00265.wav', ...}
 -->
 
## Annotation Pipeline 🔧

1. **Automatic Detection**  
   - NV detection using [BEATs](https://arxiv.org/abs/2409.09546)  
   - Emotion classification with [emotion2vec+](https://arxiv.org/abs/2402.XXX)  
   - ASR transcription via Canary model  

2. **Human Validation**  
   - 3 annotators per sample  
   - Filtered non-English/multi-speaker clips  
   - NV/emotion validation and refinement  

3. **Fusion Algorithm**  
   - Majority voting for final transcriptions  
   - Pyalign-based sequence alignment  
   - Multi-annotator hypothesis merging  


## Benchmark Results 📊


Fine-tuning CosyVoice-300M on NonverbalTTS achieves parity with state-of-the-art proprietary systems:
|Metric |	NVTTS |	CosyVoice2 |
|    ------- | ------- | -------   | 
|Speaker Similarity |	0.89 |	0.85 |
|NV Jaccard (Laugh)	| 0.92 |	0.74 |
|Human Preference | 	33.4% |	35.4% |


## Use Cases 💡
- Training expressive TTS models
- Zero-shot NV synthesis
- Emotion-aware speech generation
- Prosody modeling research

## License 📜
- Annotations: CC BY-NC-SA 4.0
- Audio: Adheres to original source licenses (VoxCeleb, Expresso)


## Citation 📝

```
@dataset{nonverbaltts2024,
  author = {Anonymous},
  title = {NonverbalTTS Dataset},
  month = December,
  year = 2024,
  publisher = {Zenodo},
  version = {1.0},
  doi = {10.5281/zenodo.15274617},
  url = {https://zenodo.org/records/15274617}
}
```