AI Now Pinpoints Where Sounds Come From — Even With Messy Microphones

Sound event localization and detection (SELD) — figuring out what a sound is and where it comes from — often relies on first-order Ambisonics (FOA) input, but getting useful FOA representations from irregular microphone arrays remains challenging. A new paper proposes a two-stage SELD framework that learns a task-oriented, FOA-compatible representation from microphone-array signals. A neural residual encoder first refines conventional FOA encoding through a signal-dependent correction, and a teacher–student scheme then transfers event and spatial knowledge from theoretical FOA representations via frame-level permutation-invariant knowledge distillation. In experiments on synthetic scenes with tetrahedral and 12-channel Benchmark arrays, along with real stationary-source recordings from the LOCATA dataset, teacher guidance consistently improved downstream SELD performance and substantially reduced localization error. Signal-level analysis further showed that lower FOA reconstruction error does not necessarily correspond to better SELD performance, indicating the distilled representation is optimized mainly for task-relevant spatial information rather than strict FOA reconstruction. The work is submitted to ICASSP 2027.