Features for the Classification and Clustering of Music in Symbolic Format

• Autor(a) principal: Bernardo, Alexandre
• Data de Publicação: 2008
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10451/13947
• topic: MIDI Track Classification; Genre Classification; Music Information Retrieval; Neural Networks; Machine Learning; Language Modeling

Music Information Retrieval is, nowadays, a highly active branch of research and development in the computer science field, and focuses several topics, including music genre classification. The work presented in this paper focus on Track and Genre Classification of music stored using MIDI format, To address the problem of MIDI track classification, we extract a set of descriptors that are used to train a classifier implemented by a Neural Network, based on the pitch levels and durations that describe each track. Tracks are classified into four classes: Melody, Harmony, Bass and Drums. In order to characterize the musical content from each track, a vector of numeric descriptors, normally known as shallow structure description, is extracted. Then they are used as inputs for the classifier which was implemented in the Matlab environment. In the Genre Classification task, two approaches are used: Language Modeling, in which a transition probabilities matrix is created for each type of track (Melody, Harmony, Bass and Drums) and also for each genre; and an approach based on Neural Networks, where a vector of numeric descriptors is extracted from each track (Melody, Harmony, Bass and Drums) and fed to a Neural Network Classifier. Six MIDI Music Corpora were assembled for the experiments, from six different genres, Blues, Country, Jazz, Metal, Punk and Rock. These genres were selected because all of them have the same base instruments, such as bass, drums, piano or guitar. Also, the genres chosen share some characteristics between them, so that the classification isn't trivial, and tests the classifiers robustness. Track Classification experiments using all descriptors and best descriptors were made, showing that using all descriptors is a wrong approach, as there are descriptors which confuse the classifier. Using carefully selected descriptors proved to be the best way to classify these MIDI tracks. Genre Classification experiments showed that the Single-Instrument Classifiers achieved the best results. Four genres achieved higher than 80% success rates: Jazz, Country, Metal and Punk. Future work includes: genetic algorithms; structurize tracks and songs; merge all presented classifiers into one full Automatic Genre Classification System