Drum Transcription via Classification of Bar-Level Rhythmic Patterns

We propose a novel method for automatic drum transcription from audio that achieves the recognition of individual drums by classifying bar-level drum patterns. Automatic drum transcription has to date been tackled by recognising individual drums or drum combinations. In high-level tasks such as audio similarity, statistics of longer rhythmic patterns have been used, reflecting that musical rhythm emerges over time. We combine these two approaches by classifying bar-level drum patterns on sub-beat quantised timbre features using support vector machines. We train the classifier using synthesised audio and carry out a series of experiments to evaluate our approach. Using six different drum kits, we show that the classifier generalises to previously unseen drum kits when trained on the other five (80% accuracy). Measures of precision and recall show that even for incorrectly classified patterns many individual drum events are correctly transcribed. Tests on 14 acoustic performances from the ENST-Drums dataset indicate that the system generalises to real-world recordings. Limited by the set of learned patterns, performance is slightly below that of a comparable method. However, we show that for rock music, the proposed method performs as well as the other method and is substantially more robust to added polyphonic accompaniment.