Time frequency domain deep CNN for automatic background classification in speech signals

Abstract Many application areas, such as background identification, predictive maintenance in industrial applications, smart home applications, assisting deaf people with their daily activities and indexing and retrieval of content-based multimedia, etc., use automatic background classification using speech signals. It is challenging to predict the background environment accurately from speech signal information. Thus, a novel synchrosqueezed wavelet transform (SWT)-based deep learning (DL) approach is proposed in this paper for automatically classifying background information embedded in speech signals. Here, SWT is incorporated to obtain the time-frequency plot from the speech signals. These time-frequency signals are then fed to a deep convolutional neural network (DCNN) to classify background information embedded in speech signals. The proposed DCNN model consists of three convolution layers, one batch-normalization layer, three max-pooling layers, one dropout layer, and one fully connected layer. The proposed method is tested using various background signals embedded in speech signals, such as airport, airplane, drone, street, babble, car, helicopter, exhibition, station, restaurant, and train sounds. According to the results, the proposed SWT-based DCNN approach has an overall classification accuracy of 97.96 (± 0.53)% to classify background information embedded in speech signals. Finally, the performance of the proposed approach is compared to the existing methods.