Detecting and Describing Percussive Events in Polyphonic Music

The present thesis deals with the automatic description of percussive events in "real world" polyphonic music. After taking a pattern recognition approach we evaluate "bag of frames" and "object-level temporal evolution" descriptors extracted from 153 frame-level features adding up a total of more than 1450 descriptors. Three binary instrument-level support vector machines models are built from a training set of more that 100 songs and 10 genres. We observe an improvement in the classification results when object-level temporal evolution descriptors are added to the feature set. Then we evaluate the binary models within a whole drum transcription system achieving comparable results with state of the art algorithms. Finally we present 17 mid-level percussion descriptors and evaluate their usefulness among MIR tasks like genre classification, danceability estimation and Western non-Western discrimination. We conclude that the presented percussion-related descriptors provide complementary information to "classic" descriptors that could help in the previously mentioned MIR tasks.