Classification of Melodies by Composer with Hidden Markov Models

In this paper, we use Hidden Markov Models (HMMs) for abstracting the style of a composer and for recognizing it out of an unknown excerpt. We employed a data set of 605 musical themes written by five well-known composers (Mozart, Beethoven, Dvorak, Stravinsky, Beatles). A preliminary investigation based on descriptive statistics served the purpose of choosing a group of suitable music representations. Then, for each representation and for each composer a HMM was trained with the subset of melodic lines extracted from his pieces. An unknown melody is then classified as belonging to a composer if the corresponding HMM gives the highest probability for that sequence. Experiments with Markov chains and tests on human subjects were used as a term of comparison. The best results achieved with HMMs was 42% of successful classifications on average, obtained with an alphabet of intervals between -10 and +10 semitones and with HMMs of order 18. In the case of human classification based only on stylistic assumption, we measured 24,6% for music amateurs and 48% for music experts. In conclusion, HMMs performed nearly as a music expert in the classification of melodies by composer; nevertheless, memory models have been proven to play a fundamental role in the process of music classification and need to be taken into consideration for practical applications.