ISCA Archive Interspeech 2023
ISCA Archive Interspeech 2023

FlexiAST: Flexibility is What AST Needs

Jiu Feng, Mehmet Hamza Erol, Joon Son Chung, Arda Senocak

The objective of this work is to give patch-size flexibility to Audio Spectrogram Transformers (AST). Recent advancements in ASTs have shown superior performance in various audio-based tasks. However, the performance of standard ASTs drastically drops when evaluated at different patch sizes than they were trained at. As a result, AST models are typically retrained to accommodate changes in patch sizes. To overcome this limitation, this paper demonstrates a training procedure to provide flexibility to standard AST models without architectural changes, allowing them to work with various patch sizes at the inference stage - FlexiAST. This proposed training approach simply utilizes random patch size selection and resizing of patch and positional embedding weights. Our experiments show that FlexiAST gives similar performance to standard AST models while maintaining its evaluation ability at various patch sizes on different datasets for audio classification tasks.