Mining digital music score collections: melody extraction and genre recognition

In the field of computer music, pattern recognition algorithms are very relevant for music information retrieval (MIR) applications. Two challenging tasks in this area is the automatic recognition of musical genre and melody extraction, having a number of applications like indexing and selecting musical databases. One of the main references for music is its melody. In a practical environment of digital music score collections the information can be found in standard MIDI file format. Music is structured as a number of tracks in this file format, usually one of them containing the melodic line, while others tracks contain the accompaniment. Finding that melody track is very useful for a number of applications, like speeding up melody matching when searching in MIDI databases, extracting motifs for musicological analysis, building music thumbnails or extracting melodic ringtones from MIDI files. In the first part of this chapter, musical content information is modeled by computing global statistical descriptors from track content. These descriptors are the input to a random forest classifier that assigns the probability of being a melodic line to each track. The track with the highest probability is then selected as the one containing the melodic line of the MIDI file. The first part of this chapter ends with a discussion on results obtained from a number of databases of different music styles. The second part of the chapter deals with the problem of classifying such melodies in a collection of music genres. A slightly different approach is used for this task, first dividing a melody track in segments of fixed length. Statistical features are extracted for each segment and used to classify them as one of several genres. The proposed methodology is presented, covering the feature extraction, feature selection, and genre classification stages. Different supervised classification methods, like Bayesian classifier and nearest neighbors are applied. As a proof of concept, the performance of such algorithms against different description models and parameters is analyzed for two particular musical genres, like jazz and classical music.This work is supported by the spanish national projects: GV06/166 and CICyT TIN2006–14932–C02, partially supported by EU ERDF and the Pascal Network of Excellence