Compression

JEPA as a Neural Tokenizer: Learning Robust Speech Representations with Density Adaptive Attention

Connecting JEPA representations to speech compression through density-adaptive attention and compact audio tokens.