Identification of Covid-19 Using Speech Recognition and Convolutional Neural Networks

This paper presents a new approach to predict possible covid-19 infected patients using convolution neural networks. We have used the open data source called Coswara dataset which consists of voice clips of healthy people and covid-19 infected people. After preprocessing the covid positive data and negative data, the sound signal features were extracted. Then, those extracted features were fed into a deep convolutional neural network and train it to predict covid-19 positiveness. Novelty of this project, compared to the similar recent research, is exploring all the possible human voices such as coughing, breathing, counting, and pronunciation which are linked with COVID-19 prediction to achieve higher accuracy. Also, understanding which human voice is mostly related to COVID-19 infection. Finally, which sound feature is mostly related to the COVID-19 infected voices. After training the convolutional neural network, we could achieve 71% of accuracy for breath deep voice signal. Also, we could achieve 68% accuracy for cough heavy signal. We could find that pronunciation of vowels is linked with COVID-19 prediction. Hence, we could conclude that covid-19 can be identified using the sound signals.