Composer Identification using PPM

Melodic and Rhythmic Pattern Analysis Using Prediction by Partial Matching (PPM)

Master ProjectThis project analyses composer identification through melodic and rhythmic pattern recognition using Markov Model. It was the main research project developed during my Master’s degree in Informatics at Universidade Federal da Paraíba, Brazil.

(Melodic and Rhythmic Pattern Analysis Using Prediction by Partial Matching)

The development of Information Theory allowed various forms of data processing that are relevant to any area of science today. Among them, we have the characterization of the self-information, where we have the possibility to determine the information given in favor of a future analysis, and, from this point, the probability of an element in a given context is to be evaluated under the entropy analysis. By this concept, various models of compressors were developed based on prediction, and the Prediction by Partial Matching algorithm (PPM) has results closest to the maximum entropy of a given input with context dependence. The PPM performs prediction by partial correlation between the elements, allowing for pattern recognition and is used in several areas.

The aim of this work is the efficiency evaluation of the use of PPM in symbolic audio files in order to be used in pattern recognition over the melodic and rhythmic patterns of melodies. The tests were conducted using melodies on MIDI files, creating models from the melodic and rhythmic parts, evaluating the efficiency of the models through cross-validation.

The results obtained with the first tests carried out on works for solo violin have been meant as possibility ways to use. There was an average hit rate of 80%, even without considering all the important concepts of Music Theory. With these results, we carried out a test of similarity of melodies, 30 melodies were sought through a bank of 5223 files with great significant results. The analysis of melodies with certain concepts using the PPM confirms the PPM as a versatile algorithm for pattern recognition in melodic sequences, considering the modeling of valid input data used in this work.

Keywords: Music analysis, PPM, MIDI

Leonardo Vidal Batisa

Final dissertation (portuguese):

DE CARVALHO JUNIOR, ANTONIO DEUSANY; BATISTA, LEONARDO VIDAL . Composer Classification in Symbolic Data Using PPM. In: 2012 Eleventh International Conference on Machine Learning and Applications (ICMLA), 2012, Boca Raton. 2012 11th International Conference on Machine Learning and Applications. p. 345-350.

Project code: PPMMDDJ