PAT: Parameter-Free Audio-Text Aligner to Boost Zero-Shot Audio Classification

Overview

PAT (Parameter-Free Audio-Text Aligner) is a simple, training-free method designed to enhance zero-shot audio classification using Audio-Language Models (ALMs), particularly those similar to CLAP. It bolsters cross-modal interactions by enriching both textual and audio representations through mutual feedback loops—without introducing new parameters or requiring additional training.