An Enhanced Approach for Environmental Sound Classification Using Multi-window Augmentation

The design and deployment of environmental sound classification (ESC) systems face substantial challenges relating to data availability and audio degradation brought on by background noises. Deep learning-based ESC systems become overfitted while modeling due to a lack of training data. In this paper, we propose an ESC model based on an enhanced multi-window spectrogram augmentation scheme using a gated convolutional neural network. The proposed method continually generates the spectrograms of the audio signal required for training, using multiple windows in terms of size (overlapping values) and type. To address the issues of overfitting and vanishing gradients, the study employs a hybrid model of convolutional neural networks and long short-term memories. To prevent overfitting, the LSTM architecture includes a dropout layer and the features extracted are provided as the input to the convolutional neural networks. The results of the experiment show that our proposed approach produces promising results for ESC problems.