Classification of Ambient Noises in Signals Using 2D Fully Convolutional Neural Network

Using a 2D convolution neural network (CNN) that develops a representation directly from the audio data, we describe a comprehensive method for ambient sound categorization in this research. Multiple convolutional layers are utilized to train a wide variety of filters appropriate to the classification job and to capture the fine-grained temporal structure of the input signal. Since the suggested method uses a sliding window to divide the signal into overlapping frames, it can handle signals of varying lengths. We compare several designs that take various input sizes into account, one of which uses a Gammatone filter bank to simulate an individual's auditory filtering response in the cochlea as the initialization for the first convolutional layer. Experimental findings reveal that the suggested end-to-end strategy to ambient sound classification achieves 89% mean accuracy, as measured on the UrbanSound8k dataset. Therefore, the suggested method achieves better outcomes than other state-of-the-art methods that rely on either manually-created features or 2D representations. The suggested method also achieves better results than any other methods that feed raw audio signals into the classifier. The suggested method also requires less data for training since its number of parameters is limited in comparison to other designs in the literature.