Chord classification of multi-instrumental music using exemplar-based sparse representation

This paper describes a chord classification system based on sparse signal representation. We aim at determining the key and type of a multi-instrumental chord from a short frame of music audio. An exemplar set consisting of magnitude spectra of single notes is constructed. Magnitude spectrum of a chord frame is decomposed into a weighted sum of note spectra by solving an ℓ₁-norm minimization problem. The weights representing the same pitch class is summed to form a pitch class vector. The chord label can be found by calculating the product index of the legitimate chords and selecting the one with the largest index. In this research, we have evaluated our system using synthesized flute and piano chords. The experimental result shows that our approach is more effective when compared with the chroma-based baseline systems.