AMBISONIC-DML: Higher-Order Ambisonic Music Dataset for Spatial AI Generation

We present AMBISONIC-DML, a dataset of 120 musical excerpts with 5th-order Ambisonic (HOA5) spatialization, aligned dry stems, and motion trajectories at 50 fps. The recordings were captured in studio conditions and spatialized post-production with composer-specified trajectories aligned to musical structure. Objective analyses confirm high-precision motion tracking and minimal spectral diffusion, while listening tests show that HOA5 enhances immersion and localization. Beyond benchmarking spatial audio, AMBISONIC-DML supports generative models that jointly learn musical content and spatial motion, providing a foundation for spatially aware music generation.