Disease Identification Using Speech Processing

Abstract

Speech is an important tool for communication. When a person is in the voiced condition the vocal cords come closer and the glottis is partially closed. The airflow which passes through glottis is disturbed by vocal cords and speech waveform are generated. The person who is suffering from the vocal cord paralysis, vocal cord blister, lungs are filled with fluid, airway blockage cannot generate a similar waveform as a healthy person. The audio feature extraction technique is to find out the difference between the healthy and pathological voice disorder. After that extracted features are pass to the machine learning to classify and predict the disease label. The ICBHI challenge 2017 and Coswara project COVID 19 datasets are used for disease prediction. Random Forest gave 90.90 %, KNN gave 80.51%, SVM gave 87.01% and XGBoost gave 85.71% accuracy on ICBHI challenge dataset. Radom forest gave 97.42%, KNN gave 89.42%, XGBoost gave 92.85% accuracy on Coswara COVID-19 dataset. As part of another experiment CNN model is used. Extracted MFCCs passed to CNN model. CNN model gave 94.06% accuracy on ICBHI challenge dataset and 70.31 % on COVID – 19 dataset. The COVID-19 detection using a screening approach is more reliable to reduce the spreading of the coronavirus. So here, a web application is developed where the user needs to record 9 audio files and the CNN base model predicted result is displayed on the screen.