Downstream Task Performance of Pre-Trained Model For Outdoor Audio Event Classification

Our surroundings’ auditory landscapes are a wealth of knowledge, providing insights into a range of outdoor pursuits. The automatic classification of these actions using audio event detection and classification (AEDC) systems has a great deal of potential applications in environmental monitoring, security, surveillance, and driverless cars. Even though AEDC has advanced significantly in indoor settings, difficulties still exist outside because of the variable and changing acoustic conditions brought on by elements like the weather, different sound sources, and ambient noise from traffic and industry. This study suggests an outdoor audio event categorization model based on convolutional neural networks (CNNs). The suggested model shows passable accuracy by utilizing adaptation to the downstream task using the ESC-50 dataset and transfer learning from a pre-trained model. The effectiveness of multi-class audio classification models in downstream tasks is analyzed in this paper, with an emphasis on the effect of an increasing number of output classes on accuracy. Three models—three, four, and five classes—with different output class configurations are used in the study, and their training and validation accuracies are assessed. Although the accuracy scores above 80% are commendable, the data show a discernible reduction in performance as the number of classes rises. Notably, the three-class model attains a validation accuracy exceeding 90%, whereas the four-class and five-class models exhibit reduced accuracies, falling below 90% and 85%, respectively.