Drum track transcription of polyphonic music using noise subspace projection

This paper presents a novel drum transcription system for polyphonic music. The use of a band-wise harmonic/noise decomposition allows the suppression of the deterministic part of the signal, which is mainly contributed by nonrhythmic instruments. The transcription is then performed on the residual noise signal, which contains most of the rhythmic information. This signal is segmented, and the events associated to each onset are classified by support vector machines (SVM) with probabilistic outputs. The features used for classification are directly extracted from the sub-band signals. An additional pre-processing stage in which the instances are reclassified using a localized model was also tested. This transcription method is evaluated on ten test sequences, each of them being performed by two drummers and being available with different mixing settings. The whole system achieves precision and recall rates of 84% for the bass drum and snare drum detection tasks.