SO-Bench: Spatial Audio Question Answering Benchmark

SO-Bench is an audio-only spatial question answering benchmark built from first-order ambisonics (FOA) spatial audio. The dataset contains FOA waveform files paired with natural-language question-answer pairs covering sound event detection, localization, spatial relations, motion, and multi-step spatial reasoning. It was created by dieKarotte and last updated on June 13, 2026.

Use Cases

• Benchmarking spatial audio question answering models based on FOA waveforms and natural language questions.
• Training models for sound event detection and localization tasks based on spatial audio data.
• Evaluating multi-step spatial reasoning capabilities in AI systems using audio-only inputs.

Strengths

• The dataset is specifically designed as a benchmark for spatial audio question answering, providing a clear evaluation target.
• Questions cover multiple complex spatial reasoning tasks, including sound event detection, localization, and motion.

Limitations

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

Provenance

Source

dieKarotte on Hugging Face

Freshness

Last updated 2026-06-13 11:04:36; freshness should be verified.