This paper deals with the Music/Speech classification problem, starting from a set of features extracted directly from compressed audio data. The proposed classification system is able to label audio sequences stored as compressed MPEG layer III files. Decoding and analyzing in a unique stage is a fundamental tool for audio streaming applications, such as real time classification. Moreover, the techniques described herein provide useful tools in the management (data tagging, summarization, etc.) of a digital music library. The adopted set of short-time features are computed from the spectral information available in the decoding stage. In this paper, we show that for the classification problem at hand this set of features is redundant and can be dramatically pruned. To this aim we used an optimization strategy based on principal component analysis and genetic algorithms. The results show a very interesting classification accuracy using just one short-time feature.