Dataset · Loie
Audiovisual YouTube sound dataset
Loie/VGGSound
Short audio clips from YouTube videos across more than 310 classes for sound recognition.
What's inside
The dataset is based on over 200,000 short audio fragments extracted from YouTube videos covering over 310 classes of real-world sounds and environmental noise. The current Hugging Face repository reflects less than 1,000 records with image modality, although the original VGGSound corpus from Oxford University is much broader.
What you can do with it
- Training sound and acoustic event classification models
- Testing audio classifiers under real-world noise conditions
- Multimodal matching of audio and video tracks
- Tasks
- Audio classification
- Size
- up to 1K rows
- Downloads this month
- 7.3K
- Likes
- 58
Models trained on it
- DAMO-NLP-SG/VideoLLaMA2.1-7B-AVVisual question answering 509
- Hguimaraes/biome_edge_bioEmbeddings 25
- lym0302/VideoLLaMA2.1-7B-AV-CoTVisual question answering 18
- Hguimaraes/biome_small_bioEmbeddings 16
- ai-forever/kandinsky-4-v2a 13
- Hguimaraes/biome_base_bioEmbeddings 13
- lym0302/VideoLLaMA2.1-7B-AV-QAVisual question answering 9
- Cactooz/DeepMMAudioVideo text → text 0
How to load it
from datasets import load_dataset
ds = load_dataset("Loie/VGGSound")