Sound Event Detection in Constrained Real-world Environments using Fine Tuned CNNs

This study investigates the improvement in recognizing and categorizing acoustic events in complex audio environments through a combination of convolutional neural networks (CNN), log-mel spectrograms, and long-short-term memory (LSTM) networks. By leveraging these algorithms, contextual information and temporal dependencies are captured, facilitating the extraction of time-frequency features from audio data. Various deep learning models were evaluated, with the proposed CNN+LSTM architecture emerging as the top performer. Achieving an accuracy of 91% and 88% for one-fold validation and 10-fold cross-validation, respectively, on the UrbanSound8k dataset, underscores the efficacy of our approach compared to recent techniques in sound-event detection. Notably, the CNN+LSTM model offers a lightweight alternative to RNNs and pre-trained models, holding considerable promise for applications in environmental monitoring and audio surveillance domains.