This paper presents a novel method for the automatic detection of singing voice in polyphonic music recordings, that involves the extraction of harmonic sounds from the audio mixture and their classification. After being separated, sounds can be better characterized by computing features that are otherwise obscured in the mixture. A set of descriptors of typical pitch fluctuations of the singing voice is proposed, that is combined with classical spectral timbre features. The evaluation conducted shows the usefulness of the proposed pitch features and indicates that the approach is a promising alternative for tackling the problem, in particular for not much dense polyphonies where singing voice can be correctly tracked. As an outcome of this work an automatic singing voice separation system is obtained with encouraging results.