Automated composer recognition for multi-voice piano compositions using rhythmic features, n-grams and modified cortical algorithms

Abstract

With the explosive growth of digital music data being stored and easily reachable on the cloud, as well as the increased interest in affective and cognitive computing, identifying composers based on their musical work is an interesting challenge for machine learning and artificial intelligence to explore. Capturing style and recognizing music composers have always been perceived reserved for trained musical ears. While there have been many researchers targeting music genre classification for improved recommendation systems and listener experience, few works have addressed automatic recognition of classical piano composers as proposed in this paper. This paper discusses the applicability of n-grams on MIDI music scores coupled with rhythmic features for feature extraction specifically of multi-voice scores. In addition, cortical algorithms (CA) are adapted to reduce the large feature set obtained as well as to efficiently identify composers in a supervised manner. When used to classify unknown composers and capture different styles, our proposed approach achieved a recognition rate of 94.4% on a home grown database of 1197 pieces with only 0.1% of the 231,542 generated features—which motivates follow-on research. The retained most significant features, indeed, provided interesting conclusions on capturing music style of piano composers.