Acoustic Scene Classification Using Convolutional Neural Network

Abstract

This proposed research work presents acoustic scene classification (ASC) which is an errand to relate a semantic name to a sound stream that distinguishes the environment in which it has been delivered. ASC can be applied in many areas including mobile robot navigation systems and context-aware devices, such as an automatically mode-switching smart phones according to the current acoustic environment. Proposing a strong ASC system is difficult because the sound from natural setting compromises numerous audio sources and also the microphones do not seem to be organized in a very controlled condition. Furthermore, not all sounds from long-duration audio data are relevant for identifying scene label. The dataset for this assignment is that the DCASE 2018 dataset collected from Tampere University of Technology, comprising of sound recordings from different scenes like airport, metro station, shopping mall, etc. For each location, there are 5–6 min of audio files. We propose to implement the ASC task using convolutional neural network (CNN) that performs the task of classification. The audio files are converted to log mel-spectrograms which are provided as input to CNN. Upon training the CNN model by varying the number of layers and the hyperparameters, it is observed that significant accuracy of 78.4 and 73.84% has been achieved for the inputs RGB scale spectrograms and grayscale spectrograms, respectively.