The idea presented in the paper consists in considering, instead of (or in addition to) direct measurements, the correlation of the measurements between pairs of vowels for speech and speaker recognition. The correlation is defined as the F-ratios in ANOVA, i.e. ratios of between – to within – speech vector parameters variation In speech recognition, the philosophy of determining significant features that are independent of the speaker centers on the search for large values of F (small correlation). For speaker recognition purposes we intend to find the speaker dependent features coefficients for which the values of F have typical large dispersion. The pilot experiments showed the this idea is suitable as the supplementary method for speaker recognition. In the speech recognition domain rather than looking for universal features in all the vowels, pair of vowel classes were subjected to analysis. Produced by 10 different speakers, the 15 pairs of the Polish vowel classes were defined in terms of cepstral coefficients, assumed to be the most useful in a dichotomous classification. The selected criterion was the value of the correlation of the respective cepstral coefficients for each pair of vowel types.