VGGsound: Audio-Visual Dataset for Natural Sounds
Audio-visual dataset containing clips of natural sounds extracted from YouTube videos, covering a wide variety of sound environments.
Over 200,000 video/audio clips, 10-second segments, 310+ sound classes, CSV files for metadata
Creative Commons Attribution 4.0 International License (CC BY 4.0)
Description
VggSound is a multimodal dataset combining audio and video clips from YouTube videos. It includes over 200,000 10-second clips, covering over 310 varied sound classes in natural and noisy environments. Metadata is provided as CSV files including YouTube URLs, timprints, labels, and train/test scores.
What is this dataset for?
- Training sound recognition models in real contexts
- Develop multimodal audio-video analysis systems
- Evaluating deep learning architectures on data “in the wild”
Can it be enriched or improved?
Yes, the dataset can be enriched by adding additional annotations, transcripts, or more accurate labels. The duration of the clips also allows cutting or recombining for specific tasks.
🔎 In summary
🧠 Recommended for
- Audio-visual researchers
- Sound recognition system developers
- Multimodal projects
🔧 Compatible tools
- PyTorch
- TensorFlow
- Hugging Face Datasets
- Librosa
- OpenAI Whisper
- MMF
💡 Tip
Pre-process videos with robust audio/video extractors to ensure the quality of the input data.
Frequently Asked Questions
Does this dataset include video files or only YouTube links?
The dataset mainly provides YouTube and timestamp links, you have to extract the clips using this information.
What is the average length of audio/video clips?
Each sample lasts about 10 seconds, which is enough to learn sound characteristics.
Can this dataset be used for commercial projects?
Yes, the CC BY 4.0 license allows commercial use under the condition of correct attribution.
