Deep learning and machine learning-based voice analysis for the detection of COVID-19: A proposal and comparison of architectures

Highlights：

• Pulmonary pathologies can be uniquely detectable from the study of the voice signal.

• Current screening techniques for COVID-19 are limited in both accuracy and frequency in time.

• Custom Adaboost and CNN architectures are employed and compared for the detection of COVID-19 from smartphone recordings.

• Acoustic features are identified as voice biomarkers for COVID-19; the RASTA filtering is a noise-robust, effective domain.

• COVID-positive and recovered subjects can be discriminated from healthy subjects.

摘要

•Pulmonary pathologies can be uniquely detectable from the study of the voice signal.•Current screening techniques for COVID-19 are limited in both accuracy and frequency in time.•Custom Adaboost and CNN architectures are employed and compared for the detection of COVID-19 from smartphone recordings.•Acoustic features are identified as voice biomarkers for COVID-19; the RASTA filtering is a noise-robust, effective domain.•COVID-positive and recovered subjects can be discriminated from healthy subjects.

论文关键词：ML,Machine Learning,CNN,Convolutional Neural Network,DL,Deep Learning,MFCC,Mel-frequency Cepstral Coefficients,P,Positive subjects,R,Recovered subjects,H,Healthy control subjects,NS,Nasal Swab,PCR,Polymerase Chain Reaction-based molecular swabs,1E,Vowel /e/ vocal task,2S,Sentence vocal task,3C,Cough vocal task,PvsH,Positive versus Healthy subjects comparison,RvsH,Recovered versus Healthy subjects comparison,SVM,Support Vector Machine,CFS,Correlation-based Feature Selection,RF,Random Forest,ReLu,Rectified Linear Unit,ROC,Receiver-Operating Curve,COVID-19,Speech processing,Classification,Deep learning,Adaboost

论文评审过程：Received 3 December 2021, Revised 18 June 2022, Accepted 22 July 2022, Available online 28 July 2022, Version of Record 11 August 2022.