In this paper we investigate the use of the time-frequency representation of speech signals, and more specifically of the second- and third-order Wigner distributions, for the hierarchical classification of speech phonemes. We focus on the obstruent rather than the sonorant class of speech phonemes. Our aim is to obtain a new set of recognition features, that are both well-suited to the non-stationary nature of these sounds and robust to noise present when speech is produced under real conditions. A three-level hierarchical classification scheme for obstruent sounds is implemented and tested on real speech data obtained from TIMIT/DARPA, either noise-free or with additive noise, recorded in the interior of a moving car. Satisfactory classification scores are obtained at all three levels, from both distributions; the advantage of third-order over second-order Wigner, however, is marginal, not justifying the extra computations required.