I took the Tensorflow RNN implementation from https://github.com/sherjilozair/char-rnn-tensorflow and modified it to generate MIDI-files.
The LSTM learns a next-character probability distribution from text files. Given a starting string, it can then generate text following the learned text-style.
To make this work for MIDI-files, I translated the MIDI-messages to "chars".
Note-on-events are coded to numbers according to their note-value 0-127, Note-off-events to 128-255.
To encode the time between two messages, I used the intervalls 1, 5, 15, 30, 60, 120, 240, 480, 960 MIDI-ticks and coded them to the
numbers 256-264. Finally the end-of-file-message is coded to 265.
Velocity and track-nrs are ignored for now to keep it simple.
Training data:
113 ragtime piano songs all transposed to C-major or to its parallel key a-minor.
The main modification to the original RNN-implementation is the way the samples are generated. I tried 4 modes to choose the next message:
- The maximum probability.
- If the maximum is a note-on then make a pick weighted by all probability values.
- If the maximum is a note-on then choose another note-on by a weighted pick according to all note-on probability values.
- If (iterator index) % (some handtuned value around 20) == 0 then choose mode 3 else choose mode 1.
Mode 4 yielded the best results. I've included 3 examples to listen to:
song-wo-3-3-long.mid
song-wo-3-4-solo1.mid
song-wo-3-5-solo2.mid
Best-of video:
