Sources of variability in the speech signal and their statistical treatment are discussed. In the experiment, a total of 3390 vowel spectra obtained from a specially constructed 100-word list read by five male and five female speakers were measured, described in terms of five variables, F1F2, F3, F4, F5 and subjected to statistical analysis. First, the 5D data are normalized, and then treated using a novel discriminant analysis based on a multi-stage comparison of pairs of classes (vowel phonemes) as represented in the data. The data of the five male and the five female voices are analyzed as two separate samples and the results are compared with those obtained using classical discriminant analysis.