A study of the characteristics of wheeze, cough, and speech sounds
Ng, Celestine RuiTing
Date of Issue2018
School of Electrical and Electronic Engineering
Over the years, many studies have developed signal processing algorithms to automatically detect the presence of wheeze in a controlled environment. However, little research was done in a practical environment where a mixture of other sounds such as cough and speech could be present. The objective of this project was to study the characteristics of sounds produced by wheeze, cough and speech. Feature extraction methods used for this research was Mel-Frequency Cepstral Coefficients (20 MFCCs), Entropy Mean, Entropy Difference and Entropy Ratio. Out of the 23 features extracted by the mentioned proposed feature extraction algorithms, 2 best features were selected using Fisher’s Ratio method with a threshold value of 0.5. Characterisation was done using a 2-Dimensional (2D) representation and data samples were modelled in rectangular shape, then further enhanced to a rotated ellipse. A validation test was carried out to measure the overall characterisation accuracy for the elliptic model which yielded a percentage of approximately 92.2%. The project was further improved beyond its scope, and a simple classification method was used. This was done by plotting linear lines to distinctively separate the classes of sounds into different regions. The accuracy of the classification method used in this project was approximately 88.84%.
Final Year Project (FYP)
Nanyang Technological University