Automatic music classification aims at grouping unknown songs in predefined categories such as music genre or induced emotion. To obtain perceptually relevant results, it is needed to design appropriate features that carry important information for semantic inference. In this paper, we explore novel features and evaluate them in a task of music automatic tagging. The proposed features span various aspects of the music: timbre, textual metadata, visual descriptors of cover art, and features characterizing the lyrics of sung music. The merit of these novel features is then evaluated using a classification system based on a boosting algorithm on binary decision trees. Their effectiveness for the task at hand is discussed with reference to the very common Mel Frequency Cepstral Coefficients features. We show that some of these features alone bring useful information, and that the classification system takes great advantage of a description covering such diverse aspects of songs.