SeqFusionNet: A hybrid model for sequence-aware and globally integrated acoustic representation

Abstract

Animals communicate information primarily via their calls, and directly using their vocalizations proves essential for executing species conservation and tracking biodiversity. Conventional visual approaches are frequently limited by distance and surroundings, while call-based monitoring concentrates solely on the animals themselves, proving more effective and straightforward than visual techniques. This paper introduces an animal sound classification model named SeqFusionNet, integrating the sequential encoding of Transformer with the global perception of MLP to achieve robust global feature extraction. Research involved compiling and organizing four common acoustic datasets (pig, bird, urbansound, and marine mammal), with extensive experiments exploring the applicability of vocal features across species and the model’s recognition capabilities. Experimental results validate SeqFusionNet’s efficacy in classifying animal calls: it identifies four pig call types at 95.00% accuracy, nine and six bird categories at 94.52% and 95.24% respectively, fifteen and eleven marine mammal types reaching 96.43% and 97.50% accuracy, while attaining 94.39% accuracy on ten urban sound categories. Comparative analysis shows our method surpasses existing approaches. Beyond matching reference models on UrbanSound8K, SeqFusionNet demonstrates strong robustness and generalization across species. This research offers an expandable, efficient framework for automated bioacoustic monitoring, supporting wildlife preservation, ecological studies, and environmental sound analysis applications.