Semi-Supervised Acoustic Scene Classification with Test-Time Adaptation

Acoustic Scene Classification (ASC) plays a crucial role in audio signal processing, with applications ranging from urban soundscapes to smart homes. However, challenges like domain shift and scarce labeled data hinder its development, highlighting the need for semi-supervised learning strategies. In the context of ICME 2024 Grand Challenge, aimed at the semi-supervised acoustic scenes classification under domain shift, our endeavor has been to devise a system that navigates these challenges. Our submission outlines a semi-supervised ASC system that employs pretraining on available datasets, followed by finetuning through FixMatch and pseudo-labeling, and concludes with test-time adaptation. This approach seeks to effectively utilize unlabeled data and mitigate domain shift, ultimately enhancing the ASC system's performance. Our final entry achieved a third-place position with a macro accuracy rate of 70.0% on the evaluation set.