Acoustic scene classification using self-determination convolutional neural network

A practical acoustic scene classification (ASC) system identifies the scene of environment by analyzing the properties of audio data that characterize the surroundings. Proposing a robust ASC system is challenging as the sound from natural environment compromises various audio sources and the microphones are not arranged in a controlled condition. Furthermore, not all sounds from long-duration audio data are relevant for identifying scene label. Some sounds may contribute to the noises or the mix of various scene labels. In this work, the problem of mixed information from various sources in natural environment is addressed by proposing a classification system that adapts to the data for each frames of long-duration audio data. The proposed ASC system is designed based on deep learning based approach, and is motivated by the researches in computer vision field. Our method is evaluated on the TUT acoustic scenes 2016 dataset. Several ASC systems have been implemented and discussed in the experiments. The results show the superiority of proposed system versus another systems that have been discussed in this work.