AudioPG trains top-performing audio models in 20 mins using only synthetic sounds

AudioPG (Procedural Generation for Audio) is a new paradigm for self-supervised audio representation learning that completely eliminates the need for real-world audio recordings during pre-training. Instead, it generates training data on-the-fly using basic acoustic primitives (e.g., sine waves, noise, filters) and composition rules based on physical models of sound propagation. The framework uses a Transformer-based masked autoencoder that learns to reconstruct masked patches of these synthetic spectrograms. Despite being trained only on artificial sounds, the learned representations transfer effectively to real-world benchmarks, achieving state-of-the-art results: 90.60% accuracy on ESC-50 environmental sound classification, 88.17% on UrbanSound8K, and 97.03% on Speech Commands V2. A standout result is 0.546 mean average precision on FSD50K, which rivals models trained on massive real audio corpora.

Pre-training completes in under 20 minutes on a single GPU, a dramatic efficiency gain compared to traditional approaches that require hundreds of GPU-hours and large curated datasets. This speed opens up audio AI research to teams without access to massive compute or proprietary data. The paper also provides interpretability insights: latent space analysis reveals that physical factors such as fundamental frequency and relative intensity naturally separate into orthogonal subspaces, meaning the model's internal representations are linearly decodable. This makes AudioPG not only efficient but also transparent—a rare combination in deep learning. Accepted at ACM ICMR 2026, the work challenges the data-centric assumption that more real data is always better, and suggests that procedurally generated physics-based signals can serve as a powerful pre-training signal when large-scale corpora are unavailable or privacy-sensitive.