Wheeze Detection Using Cepstral Analysis in Gaussian Mixture Models

Traditional wheezes detection methods are based on the frequency and durations of acoustic signal or the location of peaks from successive spectra. In these methods, the discriminative threshold used to identify peaks usually is fixed empirically. Therefore, accuracy of detected wheeze is affected by environment noise and artificial factors. The objective of this study is to classify normal and abnormal (wheezing) respiratory sounds using cepstral analysis in Gaussian Mixture Models. The sound signal is divided in overlapped segments, which are characterized by a reduced dimension feature vectors using Mel-Frequency cepstral coefficients. In this study the ´speaker´is wheeze. During the test phase, an unknown sound is compared to all the GMM models and the classification decision is based on the Maximum Likelihood criterion. In these processes, identification is based on threshold value. If the threshold is bigger than zero, the sound is normal. Otherwise, the sound is wheeze. From experimental results, when the Gaussian mix number is 16, the accuracy of identification of wheeze is up to 90%.