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

Seungryeol Paik, Kyogu Lee

Keywords: Spatial Audio, Ambisonic, AI Music, Dataset, Higher Order Ambisonics, Spatial AI Generation, Music, Music Dataset

TL;DR: A 120-excerpt HOA5 music dataset with aligned dry stems and 50 fps motion trajectories, objectively and perceptually validated, enabling benchmarking and spatially aware generative music modeling.

Abstract: We present AMBISONIC-DML, a dataset of 120 musical excerpts rendered in fifth-order Ambisonics (HOA5) with synchronized motion trajectories sampled at 50,fps. Despite its compact size, the dataset offers high informational density through 36-channel HOA encoding, 50,fps motion capture, and structured stem-level annotations, providing the first open and reproducible resource for dynamic Ambisonic music. The dataset was recorded under controlled studio conditions with composer-defined motion aligned to phrasing and rhythm. Objective and perceptual analyses confirm accurate HOA5 encoding, balanced spatial energy, and perceptual improvements in localization and immersion. AMBISONIC-DML enables reproducible research on spatial signal processing and generative modeling.

Track: Paper Track

Confirmation: Paper Track: I confirm that I have followed the formatting guideline and anonymized my submission.

Submission Number: 23