Implementation of Audio Event Recognition for The Elderly Home Support Using Convolutional Neural Networks

Abstract

This paper proposes a development of a smart home system for assisting elderly people by implementing an Audio Event Recognition (AER). By listening to the sound in the environment, the AER recognizes audio events that have been trained and then produces a useful information. There are four pretrained audio events namely door knock, can dropped, kettle sound, and rain sound. The audio in the environment is sampled for 5 seconds. Then, the sampled audio is processed into a spectrogram with a size of 128 x 76 pixels. The spectrogram serves as an input image for the Convolutional Neural Networks (CNN) to be recognized. Finally, after the spectrogram is recognized, the system produces information and transmit it to the cloud to be gathered by a smartphone. The system was implemented using Raspberry Pi 4. The experimental results show an accuracy rate of 97.5 % and 85% with a background noise of less than 40 dB and around 40 - 60 dB, respectively.