TACO: Training-free Sound Prompted Segmentation via Semantically Constrained Audio-visual CO-factorization [2412.01488]