From Physics to Representation: Audio Learning with Synthetic Pre-training via Procedural Generation
Researchers have developed AudioPG, a novel framework for pre-training audio models using procedurally generated synthetic data instead of real-world recordings. This approach significantly reduces training costs, curation efforts, and privacy concerns. The Transformer-based model trained with AudioPG demonstrates strong performance on various real audio benchmarks, achieving high accuracy rates and completing pre-training in under 20 minutes on a single GPU. Analysis of the model's latent space reveals that physical acoustic factors emerge in distinct subspaces, leading to interpretable representations. AI
IMPACT Procedural synthesis offers an efficient and interpretable alternative for audio model pre-training, potentially reducing reliance on large real-world datasets.