Towards Large-Scale Self-Supervised Learning for Spatial Audio: Generating First-Order Ambisonics from Monophonic Recordings

Self-supervised learning (SSL) has emerged as a powerful approach for learning spatial audio representations, but its progress is constrained by the scarcity of large-scale First-Order Ambisonics (FOA) datasets. To address this limitation, we propose a fast and scalable on-the-fly spatialization framework that generates diverse FOA spatial audio samples from monophonic recordings during SSL training. Our method continuously synthesizes new spatialized samples by combining multiple monophonic audio sources with randomized spatial trajectories and overlapping sound events, significantly expanding the training data without requiring explicit labels. The spatialization process leverages room impulse responses and parametric spatialization techniques to ensure physically plausible spatial cues. By dynamically generating training data at each learning step, our approach enables large-scale SSL pre-training without the need for extensive amounts of FOA datasets. We evaluate the effectiveness of the generated dataset by pre-training a transformer-based SSL model and fine-tuning it for Sound Event Localization and Detection. Experimental results demonstrate that models trained on our synthetic dataset achieve competitive performance on real-world FOA benchmarks, bridging the gap between synthetic and real spatial audio. This work provides a scalable solution for self-supervised spatial audio learning, paving the way for advancements in spatial scene understanding, immersive media, and spatial AI applications.