Ambisonic-DML

AMBISONIC-DML is a dataset of dynamic music excerpts rendered in 5th-order Ambisonics (HOA5) with aligned dry stems and synchronized motion trajectories. The dataset is designed to support reproducible research in spatial audio signal processing and AI-based generative modeling.

The dataset consists of 120 music excerpts with musically meaningful spatial motion aligned to phrasing and rhythm. Each excerpt provides three synchronized modalities: (1) dry audio stems, (2) spatialized audio in HOA5, and (3) XYZ motion trajectories. Both per-stem HOA5 renderings and full-mix HOA5 renderings are included.

This Zenodo record serves as the canonical reference for citation and long-term preservation. It provides:- Code for evaluation and reproduction- Metadata and annotations- A complete file manifest (MANIFEST.txt)- SHA-256 checksums (SHA256SUMS)

Due to the large size of the complete dataset (approximately 634 GB), the full-resolution audio and motion files are hosted via an external read-only mirror. File integrity can be verified by matching downloaded files against the provided SHA-256 checksums.

Full dataset mirror:https://drive.google.com/drive/folders/1wLmoDqqYOStCFt_WTny7jSUxrBjV5C56?usp=sharing

The dataset is released under the Creative Commons Attribution–NonCommercial 4.0 International (CC BY-NC 4.0) license.

The dataset is associated with an accepted paper at IEEE ICASSP 2026. Citation information will be updated after the final publication.