AudioMosaic: Contrastive Audio SSL with Masked Spectrograms
Researchers introduce AudioMosaic, a contrastive self-supervised learning method for audio representation. It constructs positive pairs via structured time-frequency masking on spectrogram patches, enabling efficient large-batch training. The encoder learns discriminative utterance-level representations that transfer well across datasets and conditions, outperforming generative approaches in experiments.
Key facts
- AudioMosaic is a contrastive learning-based audio encoder for general audio understanding.
- It uses structured time-frequency masking on spectrogram patches to create positive pairs.
- The method reduces memory usage and enables efficient large-batch training.
- It learns more discriminative utterance-level representations than generative approaches.
- Representations show strong transferability across datasets, domains, and acoustic conditions.
- Extensive experiments demonstrate its effectiveness.
- The paper is available on arXiv under ID 2605.14231.
- The approach addresses challenges in contrastive audio SSL such as augmentation design and batch size.
Entities
Institutions
- arXiv