Dataset for Evaluating Sustain-Pedal Detection from Polyphonic Piano Music

To evaluate methods of sustain-pedal detection from polyphonic piano music, we built a dataset consisting of ten well-known passages of Chopin's music. Ground-truth annotations of this dataset represent sustain-pedal on/off states at every 0.1 second. This annotation was based on the sustain-pedal movement tracked by a dedicated measurement system. Music scores of the ten passages were saved in PDF files. They were performed by a pianist using a Yamaha baby grand piano situated in the studios at Queen Mary University of London. The audio were recorded at 44.1 kHz and 24 bits using the spaced-pair stereo microphone technique. A pair of Earthworks QTC40 omnidirectional condenser microphones was positioned about 50 cm above the strings. We have developed a transfer learning method such that the on/off state of the sustain pedal can be detected at every 0.1 second. The ground-truth annotation and our detection results were saved in transfer-learning-y_segment.npz, which can be loaded using numpy.load in Python. The key for passage name, ground-truth annotation and our detection results is filename_record, y_true and y_pred, respectively.