Time Variability-Based Hierarchic Recognition of Multiple Musical Instruments in Recordings

The research reported in this chapter is focused on automatic identification of musical instruments in polyphonic audio recordings. Random forests have been used as a classification tool, pre-trained as binary classifiers to indicate presence or absence of a target instrument. Feature set includes parameters describing frame-based properties of a sound. Moreover, in order to capture the patterns which emerge on the time scale, new temporal parameters are introduced to supply additional temporal information for the timbre recognition. In order to achieve higher estimation rate, we investigated a feature-driven hierarchical classification of musical instruments built using agglomerative clustering strategy. Experiments showed that the performance of classifiers based on this new classification of instruments schema is better than performance of the traditional flat classifiers, which directly estimate the instrument. Also, they outperform the classifiers based on the classical Hornbostel-Sachs schema.