TWNM FOA Benchmark: 1,000 Spatial Audio Multiple-Choice Items

1,000 multiple-choice benchmark items with first-order ambisonics audio, released in 2026. It corresponds to the evaluation set for the paper 'The World is Not Mono: Enabling Spatial Understanding in Large Audio-Language Models'. The dataset was authored by KonoyoBC and is hosted on Hugging Face.

Use Cases

• Benchmarking spatial reasoning capabilities of audio-language models based on first-order ambisonics audio.
• Evaluating model performance on multiple-choice questions derived from spatial audio scenes.
• Training models to understand and answer questions about directional and spatial properties of sound.

Strengths

• Contains 1,000 benchmark items, providing a defined scale for evaluation.
• Audio is in first-order ambisonics format, which is a specific and relevant format for spatial audio tasks.
• Directly linked to a published research paper, providing academic context.

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

KonoyoBC on Hugging Face.

Collection Method

Likely created as an evaluation benchmark for a specific research project.

Freshness

Last updated 2026-06-04 08:51:20; freshness should be verified.