This article concerns the spectral analysis of vowels sustained by hoarse and healthy speakers. Conventional cues of noise in spectral data are founded on the segregation of harmonics and inter- harmonics. What we propose here is an alternative that consists of flattening the spectral contour and performing a multivariate statistical analysis of the residual ( i.e. flattened ) spectral components. The residual spectrum is obtained by means of the wavelet transform of the power spectrum of the vowels. The results show that a principal components analysis of the flattened spectra enables hoarse voices to be separated from clear.