Data

Lakh Pianoroll Dataset

We use the cleansed version of Lakh Pianoroll Dataset (LPD). LPD contains 174,154 unique multitrack pianorolls derived from the MIDI files in the Lakh MIDI Dataset (LMD), while the cleansed version contains 21,425 pianorolls that are in 4/4 time and have been matched to distinct entries in Million Song Dataset (MSD).

Training Data

• Use symbolic timing, which discards tempo information (see here for more information)
• Discard velocity information (using binary-valued pianorolls)
• 84 possibilities for note pitch (from C1 to B7)
• Merge tracks into 5 categories: Bass, Drums, Guitar, Piano and Strings
• Consider only songs with an rock tag
• Collect musically meaningful 4-bar phrases for the temporal model by segmenting the pianorolls with structure features proposed in [1]

Hence, the size of the target output tensor is 4 (bar) × 96 (time step) × 84 (pitch) × 5 (track).

The following are two sample pianorolls seen in our training data. The tracks are (from top to bottom): Bass, Drums, Guitar, Strings, Piano.

Reference

1. Joan Serrá, Meinard Müller, Peter Grosche and Josep Ll. Arcos, “Unsupervised Detection of Music Boundaries by Time Series Structure Features,” in AAAI Conference on Artificial Intelligence (AAAI), 2012