Sonic-AM, Streaming Human Motion Generation from Spatial Audio

This paper presents Sonic-AMG, a framework for generating human motion in real-time from spatial audio. The model encodes acoustic content and body-relative source geometry as separate tokens, using a causal transformer with a flow head to generate motion latents. It introduces rehearsal training and spatial feedback to address training-inference mismatches, achieving state-of-the-art results on the SAM benchmark.