A Semi-Supervised Acoustic Scene Classification Network Based on Multi-Modal Information Fusion

This paper presents our semi-supervised acoustic scene classification (ASC) framework submitted to the APSIPA ASC 2025 Grand Challenge, which focuses on city- and timeaware ASC under limited labeled data. Our approach leverages a multi-modal network architecture that fuses audio melspectrograms with spatiotemporal metadata (city identity and timestamps) to capture dynamic acoustic scene variations across urban environments. The model employs a residual-based CNN with attention mechanisms for robust feature extraction, enhanced by multi-modal fusion. To address label scarcity, we adopt a staged semi-supervised pipeline: pre-training on TAU Urban Acoustic Scenes 2020 and CochlScene datasets with specaugment and mixup augmentations, and then iterative fine-tuning on challenge data with pseudo-labeling to expand the training set was conducted, resulting in performance improvement. Experimental results demonstrate the efficacy of our city/time-aware design and semi-supervised strategies on our validation data.