Classification of distressed sounds using CNN/C-RNN

Abstract

Safety is always the utmost priority in this world where dangers are all around. There may be incidents of snakes, falling trees and even car crashing that may endanger one life. With improvements in quality of life in Singapore, the response from emergency personnel will arrive swiftly when contacted by the victim or other people. However, imagine if it were to occur in a deserted area or in a factory where no one else is present and the victim could not obtain help from any means of communication, this project will provide the solution.

By having a trained distressed sounds classifier, distressed sounds can be detected so that the investigation team or emergency personnel can seek the victim. Integration of distressed sound detection in a sound-based surveillance system can thus be implemented at several places like factories and deserted areas to extend assistance to people who are distressed, in pain or danger [1].

Hence, this project discusses the development and usage of machine learning techniques, Convolutional Neural Network (CNN) and Convolutional-Recurrent Neural Network (CRNN) model to classify distressed sounds in Singapore’s soundscape. These distressed sounds are categorized into 4 classes: non-distressed sounds, ‘Crying’, ‘Help’, and ‘Screaming’. Furthermore, the models to be implemented are inspired by VGG [2] which is widely used in image and audio classification. In general, this report shows the process of transforming audio classification into an image classification problem where CNN and CRNN can be utilized efficiently. In the end, the performance of these networks was evaluated based on several metrics but unfortunately, they have not shown a feasible result that can be implemented in real-time. CNN and CRNN models have only scored F_β score of 0.3377 and 0.3225 respectively when beta is 2. Keyword: Audio classification, Distressed Sounds, Deep Neural Network