The frequencies and levels of the first three formants of 12 Dutch vowels were measured. The vowels were spoken by 50 male speakers in an h (vowel) t context. Statistical analysis of these formant variables confirmed thatF1andF2are the most appropriate two distinctive parameters for describing the spectral differences among the vowel sounds. Maximum likelihood regions were computed and used to classify the vowels, and a score of 71.3% correct classification in the logF1‐logF2plane was obtained (87.3% if three related pairs are grouped together). These scores rose to 78.3% and 95.2%, respectively, when a simple speaker‐dependent correction was applied. The scores are comparable with those obtained in an earlier study in which a principal‐components analysis was applied to the 1/3‐oct filter levels of the same vowel sounds [Klein, Plomp, and Pols, J. Acoust. Soc. Amer.48, 999–1009 (1970)]. From the latter data a two‐dimensional representation (“optimal plane”) equivalent to the logF1‐logF2plane could be derived. The relative merits of the two approaches are discussed. For automatic speech recognition in particular, the dimensional analysis is much more attractive than the formant analysis because it is much simpler and can be carried out in real time.