Acoustic Scene Classification with Attention-based Neural Networks

Abstract

Auditory information provides great help to human beings to recognize their surroundings and positions. However, the sound we perceived in the environment is often a mixture of many sounds that happened at the same time. Therefore, developing a system to automatically extract information from unprocessed audio provides a huge potential for human beings. For example, it can be utilized by automatic diving, multimedia searching, robots, etc. In this study, we proposed two Attention-based Neural Network models to achieve automatic acoustic scene classification systems. Both two models are powerful in extracting information on audio spectrograms and classifying them into their corresponding scene labels. We applied two acoustic scene datasets to verify our model and got the best accuracies which are 15.7% and 8.0% higher than their CNN baselines.