Audio-JEPA is an adaptation of the Joint-Embedding Predictive Architecture (JEPA) for self-supervised audio representation learning. Built upon the I-JEPA paradigm, it uses a Vision Transformer (ViT) backbone to predict latent representations of masked spectrogram patches.

JEPAs for audio representation learning

Audio jepa version 1

Project to explore LeJEPA-style SSL architecture for ASR using the LibriSpeech dataset.

No description available

Experimenting with JEPA objectives for view-invariant music audio SSL

Self-supervised learning (SSL) audio embedding framework with PyTorch Lightning + Hydra supporting Audio-JEPA, RQA-JEPA, BEST-RQ (ViT based), and BEST-RQ-2.

Self-supervised audio representation learning with a masked Joint Embedding Predictive Architecture (JEPA), analyzing representation geometry and temporal dynamics.