We propose a two-class classification scheme with a small number of features for sleepiness detection. Unlike the conventional methods that rely on the linguistics content of speech, we work with prosodic features extracted by psychoacoustic masking in spectral and temporal domain. Our features also model the variations between non-sleepy and sleepy modes in a quasi-continuum space with the help of code words learned by a bag-of-features scheme. These improve the unweighted recall rates for unseen people and minimize the language dependence. Recall rates reported based on Karolinska Sleepiness Scale (KSS) for Support Vector Machine and Learning Vector Quantization classifiers show that the developed system enable us monitoring sleepiness efficiently with a lower complexity compared to the reported benchmarking results for Sleepy Language Corpus.