SO-Dataset: Large-Scale Spatial Audio in First-Order Ambisonics

SO-Dataset is a large-scale spatial audio dataset in first-order ambisonics format. It combines simulated spatial scenes and real FOA recordings, with sound event annotations mapped into a unified 63-class taxonomy based on the FSD50k dataset. The dataset was created by dieKarotte and last updated in June 2026.

Use Cases

• Train machine learning models for spatial sound event detection based on first-order ambisonics audio.
• Benchmark audio source separation algorithms using the dataset's spatial event annotations.
• Develop and evaluate acoustic scene analysis systems based on the combined simulated and real recordings.
• Research sound event classification using the unified 63-class taxonomy.

Strengths

• Dataset is described as large-scale, suggesting substantial volume.
• Employs a unified 63-class sound event taxonomy based on the established FSD50k dataset.
• Combines both simulated spatial scenes and real FOA recordings.

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

dieKarotte on Hugging Face

Collection Method

Combines simulated spatial scenes and real FOA recordings.

Freshness

Last updated 2026-06-13 02:16:20; freshness should be verified.