Figure 1 Normal and pathological signals and their TFDs.
Related Figures (6)
Fig. 2. (a) Signal “SAFETY.” (b) ATFT reconstructed “SAFETY” signal with first 10 TF functions. Fig. 3. (a) Vowel “\ a \.” (b) TFD of vowel “\ a \.” (c) Vowel “‘\ a \” added with random noise. (d) TED of vowel “\ a \” added with noise number of lower order octaves used versus the higher order oc- taves. This is correlated with the global envelope of the signal and the phonetic composition of the signal. with the first 10 TF functions. The reconstructed signal is shown in Fig. 2(b). From the figure one can clearly see the length or the octave of the Gaussian TF functions contributing to the envelope of the signal. In order to simulate voice pathology, synthetic random noise was added to the vowel segment “\ a \’ in the word “SAFETY.” The vowel “\ a \” and noise added “\ a \” are shown in Fig. 3(a) and (c). Both the vowel and the noise added vowel were decomposed using 50 iterations and their TF dis- tributions are presented in Fig. 3(b) and (d), respectively. It can be clearly observed from these figures that the ATFT decom- position approximates the periodic structures constituting the voiced segments with octaves of longer duration (tonal compo- nents), whereas for approximating the noisy vowel, in addition to few longer duration octaves, many shorter duration octaves (high-frequency support) were needed. Thus, in a fixed number of iterations, based on the structural definition of the voiced and unvoiced components, there was a shift in the balance of the Fig.4. (a) and (b)—Octave distribution of a normal and pathological signal. (c) and (d)—Energy curve of a normal and pathological signal. (e) and (f)—Frequency distribution of a normal and pathological signal. Fig. 5. ROC curves for Ocmax, Ocmean, and Fr. CLASSIFICATION RESULTS. METHOD: ORIGINAL—LINEAR DISCRIMINANT ANALYSIS, CROSS-VALIDATED—LINEAR DISCRIMINANT ANALYSIS WITH LEAVE-ONE-OUT METHOD, %—PERCENTAGE OF CLASSIFICATION talkers are clearly visible. To further investigate the total dis- criminative power of these features, the entire database was fed to a LDA-based classifier. Classification results are shown in Table I in the form of a confusion matrix with percentage clas- sification for normal and pathological signals. From the table it can be observed that out of 51 normal signals, 49 were correctly classified as normal, and 2 were misclassified as pathological. Similarly, out of 161 pathological signals, 149 were correctly classified as pathological and 12 were misclassified as normal. An overall classification accuracy of 93.4% was achieved, and this accuracy was verified using the leave-one-out method. The misclassified signals were analyzed in detail, and it was ob- served that 6 misclassified pathological signals were perceptu- ally similar to normal and that they could be hardly classified as pathological by an untrained listener. Fig. 6. Feature space: Ocmax versus Ocmean versus Fr.
