Most of the automated music analysis methods available in the literature rely on the representation of the music through a set of low-level audio features related to temporal and frequential properties. Identifying high-level concepts, such as music mood, from this "black-box" representation is particularly challenging. Therefore we present in this paper a novel music representation that allows gaining an in-depth understanding of the music structure. Its principle is to decompose sparsely the music into a basis of elementary audio elements, called musical words, which represent the notes played by various instruments generated through a MIDI synthesizer. From this representation, a music feature is also proposed to allow automatic music classification. Experiments driven on two music datasets have shown the effectiveness of this approach to represent accurately music signals and to allow efficient classification for the complex problem of music mood classification.