An Enhanced Approach for Automatic Sound Event Detection Using Neural Networks

Sound event detection (SED) is a process of identifying different types of sound events and their temporal boundaries. It has applications in home and surveillance settings, but is challenging due to background noise and overlapping events. In this paper, we developed a model for home environment sound classification using convolutional neural networks (CNNs). Our research was implemented with an audio dataset that was collected from Kaggle which contains 18,873 audio files categorized into 41 classes related to home environment. We extracted various features from audio samples, including mel-spectrogram, chromagram, spectral bandwidth, spectral centroid, and MFCC. These features were stacked together and used to train a CNN model. The overall accuracy of the model on the 41 classes was found to be 74%. The results demonstrate the effectiveness of the approach for accurate sound classification in home environments, with potential applications in smart home systems and environmental monitoring.