When we address speaker states like sleepiness, two partly competing interests can be observed: within both applications and engineering approaches, we aim at utmost performance in terms of classification or regression accuracy - which normally means using a very large feature vector and a brute forcing approach. The other interest is interpretation: we want to know what tells apart atypical (here: sleepy) speech from typical (here: non-sleepy) speech, i.e., their respective feature characteristics. Both interests cannot be served at the same time. In this paper, we preselect a small number of easily interpretable acoustic-prosodic features modelling spectrum and prosody, based on the literature and on the general idea of sleepiness being characterised by relaxation. Performance obtained with these single features and this small feature vector is compared with the performance obtained with a very large feature vector; moreover, we discuss to which extent the features chosen model relaxation as sleepiness characteristic.