Read the original document by opening this link in a new tab.
Table of Contents
1. Introduction
2. Approach
3. Experiments
Summary
The document discusses the need for flexibility in Audio Spectrogram Transformers (AST) and introduces FlexiAST, a training procedure to provide patch-size flexibility to AST models. It addresses the limitations of standard ASTs in adapting to different patch sizes and proposes a method to achieve flexibility without architectural changes. The experiments demonstrate that FlexiAST performs comparably to standard AST models on various datasets for audio classification tasks, showing improved flexibility at different patch sizes.