Deep Learning-Based Hazardous Sound Classification for the Hard of Hearing and Deaf

摘要

The hard of hearing or deaf can only access limited auditory information in dangerous situations. Therefore, development of a system for sensing hazardous auditory information may be of great help to them. However, such systems have focused on effective signal transduction when a hazardous sound is detected, and the classification of hazardous sounds has been less investigated. The present study was conducted to classify sounds by using Recurrent Neural Network (RNN)-based models, Convolutional Neural Network (CNN)-based models, the combination of the two models, and ensemble models prepared by combining various models. The experimental results showed that the accuracy of the 3-layer Long Short-Term Memory (LSTM) model was 97.63% and that of the ensemble model was 98.00%. As an attempt at real-life application of the developed model, a warning system was prepared by using Raspberry Pi and a vibrator.