A ground-truth experiment on melody genre recognition in absence of timbre

Abstract

Music genre or style is an important metadata for music collections and database organization. Some authors claim for the need of having ground truth studies on this particular topic, in order to compare results with them and lead to sound conclusions when analyzing software performances. When dealing with digital scores in any format, timbrical information is not always available or trustworthy so we have avoided this information in our computer models, using only melodic information. The main goal of this work is to assess the human ability for recognizing music genres in absence of timbre in order to assess comparatively the performance of computer models for this task. For this, we have experimented with fragments of melodies in absence of accompaniment and timbre, as our computer models do. For this particular paper we have worked with two well-stablished genres in the music literature, like classical and jazz music. A number of analyses in terms of age, group, education, and music studies of the people subjected to the tests have been performed. The results show that, on average, the error rate was about 18%. This value shows the base line to be improved for computer systems in this task without using timbrical information.