Towards Automatic Transcription of Australian Aboriginal Music

We describe a system designed for automatic extraction and segmentation of didjeridu and clapsticks from certain styles of traditional Aboriginal Australian music. For didjeridu, we locate the start of notes using a complexdomain note onset detection algorithm, and use the detected onsets as cues for determining the harmonic series of sinusoids belonging to the didjeridu. The harmonic series is hypothesised, based on prior knowledge of the fundamental frequency of the didjeridu, and the most likely hypothesis is assumed. For clapsticks, we use independent subspace analysis to split the signal into harmonic and percussive components, followed by classification of the independent components. Finally, we identify areas in which the system can be enhanced to improve accuracy and also to extract a wider range of musically-relevant features. These include algorithms such as high frequency content techniques, and also computing the morphology of the didjeridu.