Figure 1: Left: A scene with 1815 mobile sound sources. Audio is rendered in realtime with our progressive lossy processing technique using 15% of the frequency coefficients and with an average of 12 clusters for 3D audio processing. Degradations compared to the reference solution are minimal. Right: a scene with 1004 mobile sound sources, running with 25% of the frequency coefficients and 12 clusters. AbstractDespite recent advances, including sound source clustering and perceptual auditory masking, high quality rendering of complex virtual scenes with thousands of sound sources remains a challenge. Two major bottlenecks appear as the scene complexity increases: the cost of clustering itself, and the cost of pre-mixing source signals within each cluster.In this paper, we first propose an improved hierarchical clustering algorithm that remains efficient for large numbers of sources and clusters while providing progressive refinement capabilities. We then present a lossy pre-mixing method based on a progressive representation of the input audio signals and the perceptual importance of each sound source. Our quality evaluation user tests indicate that the recently introduced audio saliency map is inappropriate for this task. Consequently we propose a "pinnacle", loudness-based metric, which gives the best results for a variety of target computing budgets. We also performed a perceptual pilot study which indicates that in audio-visual environments, it is better to allocate more clusters to visible sound sources. We propose a new clustering metric using this result. As a result of these three solutions, our system can provide high quality rendering of thousands of 3D-sound sources on a "gamer-style" PC.