Soundscape Bench: 200 Multilingual Soundscapes for Universal Audio Annotation

200 held-out multilingual soundscapes provide exact, automatically-gradable answer keys for evaluating universal audio annotation. This benchmark is designed for the LAION Universal Audio Annotation Pipeline (UAAP), which describes everything audible in a clip as a structured JSON list. Each clip is built by gluing together pieces, as noted in the full description.

Use Cases

• Benchmarking audio annotation models based on the structured JSON answer keys
• Evaluating multimodal AI systems on the task of describing everything audible in a clip
• Training models for universal audio annotation based on the multilingual soundscape examples

Strengths

• 200 held-out multilingual soundscapes provide a defined evaluation scale
• Exact, automatically-gradable answer keys enable objective benchmarking
• Structured JSON list format supports the specific task of universal audio annotation

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

laion

Collection Method

Clips are built by gluing together pieces, as noted in the description.

Freshness

Last updated 2026-06-11 12:21:46; freshness should be verified