Multivariate statistical analysis of flat vowel spectra with a view to characterizing dysphonic voices

The aim of this article is to show how dysphonic voices can be characterized by means of a multivariate statistical analysis of flat vowel spectra. The spectral contour was obtained by means of a wavelet transform of the logarithmic magnitude spectrum, which was subsequently flattened to remove interspeaker variability related to the excitation and vocal tract filter functions. The results of the statistical analysis of flat spectra were the following. Firstly, principal components analysis produced markers that separated noisy from clean spectra. Secondly, the heuristic search for harmonic peaks or interharmonic dips could be omitted. Thirdly, conventional spectral markers of noise appeared as special instances of the markers that were derived statistically. Fourthly, the levels of visually assigned hoarseness and the first two principal components were significantly correlated. The assignment of different levels of (visual) hoarseness to different vowel timbres could be explained by the variability associated with the spectral contour.