This repository contains the model code and pretrained weights for "Physics-Informed Audio-Geometry-Grid Representation Learning for Universal Sound Source Localization" [1].
Overall architecture of Audio-Geometry-Grid Representation Learning (AGG-RL) framework with learnable non-uniform discrete Fourier transform (LNuDFT) layer and relative microphone positional encoding (rMPE).
AGG-RL enables universal sound source localization that generalizes to unseen microphone array geometries and customizable spherical grids.
You can infer the code by
python inference.py
Example with 13 channels and 1 speaker.
Example with 4 channels and 2 speakers.
[1] Min-Sang Baek, Gyeong-Su Kim, Donghyun Kim and Joon-Hyuk Chang "Physics-Informed Audio-Geometry-Grid Representation Learning for Universal Sound Source Localization," in Proc. International Conference on Learning Representations (ICLR), 2026, 147447–147473. [Paper]