Rethinking Procedural Audio Pre-training: Source Scaling and Objective Adaptation

The paper investigates how procedural audio should be scaled for effective pre‑training and whether training strategies from natural audio transfer to procedural data. By separating scale into formula‑class coverage and within‑class rendering diversity, the authors show that each type of scale benefits different learning formulations and downstream tasks. Their experiments reveal that procedural audio prefers lower mask ratios, and that it exhibits lower patch diversity and stronger temporal predictability compared to natural audio, leading to a proposal for source‑aware procedural pre‑training.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.