New SARL Benchmark Reveals Systematic Biases in Spatial Audio AI Models
Researchers probe how AI hears space—source factors easy, room acoustics hard.
Deep Dive
The Spatial Audio Representation Learning (SARL) benchmark is introduced for evaluating spatial information in pretrained audio models. SARL probes source-level factors (azimuth, elevation, distance, class) and room-level factors (RT60, volume, shape). Experiments show source factors are consistently easier to decode than room factors, and input configuration and training paradigm shape spatial encoding. SARL is released as an open-source benchmark for reproducible evaluation of spatial audio representations.
Key Points
- SARL evaluates both source-level factors (azimuth, elevation, distance, class) and room-level factors (RT60, volume, shape) across diverse pretrained encoders.
- Source factors are consistently easier to decode than room factors, revealing a systematic bias in current spatial audio representations.
- Input configuration and training paradigm drastically shape spatial encoding, with heterogeneous responses to controlled perturbations.
Why It Matters
Better spatial audio understanding is critical for AR/VR, hearing aids, and autonomous vehicle sound localization.