Third International Workshop on Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA 2003)
The classification performance of an automatic classifier of voice pathology for the detection of normal and pathologic voice types is presented. The proposed classification system is nonintrusive and fully automated. Speech files of sustained phonation of the vowel sound /a/ in the 'Disordered Voice Database Model 4337' provided 631 subjects of both genders (58 normal, 573 pathologic). This database includes features extracted by the Multi Dimensional Voice Program (MDVP). Mel frequency cepstral coefficients (MFCC) were extracted for all of the speech files. Discrete Fourier transform (DFT) features, Log DFT and Cepstral features were also extracted. Cross-fold validation was used to measure the classifier performance. Linear discriminant analysis was employed as the classifier model. The MDVP feature set of shimmer and signal-to-noise ratios are shown to have similar classification performance to the Log DFT and the MFCC features.
Index Terms. Voice Pathology, speech analysis, Linear Discriminant Analysis
Full Paper (reprinted with permission from Firenze University Press)
Bibliographic reference. Maguire, C. / Chazal, P. de / Reilly, R. B. / Lacy, P. D. (2003): "Identification of voice pathology using automated speech analysis", In MAVEBA-2003, 259-262.