PRISM: Training-Free Test-Time Adaptation for Audio Models

Ashish Anand Shukla · hf · 2026-08-18

PRISM is a fast, training-free test-time adaptation method addressing severe acoustic shifts in audio-text models. It reverses low-rank affine noise distortions using frozen text prototypes and geometric corrections, enhancing performance without retraining.

Original post →

More from Research

Research channel →