An Automatically Transcribed Piano Performer Dataset

Data-driven approaches for performers' style analysis need large corpora of music performances to derive expressive performance parameters. One good reason for Deep Neural Networks (DNNs) not being used for performer identification is the lack of large-scale datasets with overlapping performances by different performers. Hence, to bridge this gap, we created a score-aligned automatically transcribed performer dataset. There are a total of 474 performances in the dataset, played by 6 pianists, spanning 35 movements by 2 composers for a total of 474 Western classical piano recordings in MIDI format. Every single MIDI file that's included in this collection was derived from a piano transcription of an already-existing audio recording of a piano performance. Score files in musicXML format and the alignment results are also included in the dataset.