VGG-MonoAudio: 10,000 Triplets for Selective Video-to-Audio Generation

VGG-MonoAudio is an evaluation benchmark for text-conditioned selective video-to-audio (V2A) generation containing between 1,000 and 10,000 data triplets. Developed by jnwnlee and associated with Arxiv paper 2512.02650, the collection features synchronized video, text descriptions, and isolated audio tracks. It was constructed by synthetically mixing single-source clips from the VGGSound and UnAV-100 datasets.

Use Cases

• Evaluating text-conditioned V2A models by using the text prompt to generate audio matching a specific visual source
• Training audio source separation algorithms using the mixed-source video and the isolated target audio track
• Benchmarking cross-modal alignment between the text description and the isolated audio component

Strengths

• 1,000 to 10,000 curated triplets
• Clean isolated target audio labels for multi-source visual scenes
• Derived from established VGGSound and UnAV-100 sources

Limitations

• Visuals are synthetically mixed from single-source clips rather than naturally occurring multi-source scenes
• Limited to mono-audio output
• Small sample size compared to large-scale audio-visual pre-training sets

Provenance

Source

jnwnlee (Arxiv 2512.02650)

Collection Method

synthetic mixing of single-source clips from VGGSound and UnAV-100

Freshness

Last updated March 2026.

Geography

Global