This paper deals with the following question: can the conventionnal classification algorithms be transposed to the "hyperbaric" speech signal case? An objective estimate of discriminating features computed on synthetic signals, that is on tests signals, appears to be a relevant approach. As it is, five basic features commonly involved in voiced-unvoiced decision rules are consistently computed, first on "air" vowels signals considered as references, then on similar "heliox" vowels signals. Globally, the discriminating features abilities are preserved and are even likely to be improved, provided that thresholds are adapted.

Keywords: "hyperbaric" speech, classification features, speech modelization.