Abstract
We have previously proposed a structured sparse approach to piano transcription with promising results recorded on a challenging dataset. The approach taken was measured in terms of both frame-based and onset-based metrics. Close inspection of the results revealed problems in capturing frames displaying low-energy of a given note, for example in sustained notes. Further problems were also noticed in the onset detection, where for many notes seen to be active in the output trancription an onset was not detected. A brief description of the approach is given here, and further analysis of the system is given by considering an oracle transcription, derived from the ground truth piano roll and the given dictionary of spectral template atoms, which gives a clearer indication of the problems which need to be overcome in order to improve the proposed approach.