Classification of Environmental and Urban Sounds Using Deep Learning Techniques

Environmental sound classification (ESC) is an important task in machine learning, with broad applications ranging from urban monitoring to wildlife protection and smart cities. By enabling machines to recognize and categorize sounds from the environment, ESC systems help automate sound detection and recognition. ESC systems often struggle with noise and the diverse nature of sounds, existing methods rely on complex feature extraction processes sometimes including multiple intermediate stages. This paper proposes a transfer learning approach using the Audio Spectrogram Transformer (AST), an attention-based transformer network for audio classification. Proposed method is to combine it with a Sequential block containing LSTM, Linear, and Dropout layers. Unlike other approaches that involve multiple feature extractions, AST can directly convert the spectrogram images of audio into its features, similar to the Vision-Transformer (ViT). The model was evaluated on ESC-10, ESC-50, and UrbanSound8k datasets. The results confirm that combining transformers and LSTM shows better results in sound classification.