Recurrent Neural Network for generative MIDI music
A GRU1-based RNN2 for rhythmic pattern generation.
The RNN model is a
char-rnn
that gets trained on an input MIDI file encoded as a sequence of
unit vectors
Run $ wolframscript -configure and set the variable WOLFRAMSCRIPT_KERNELPATH to your local WolframKernel address
Run $ wolframscript -f encodeAndTrain.wl
Type the input filename5 *.mid
The trained net and decoding parameters are saved in data/.
Run $ wolframscript -f generateAndDecode.wl
Generated *.mid is saved in data/.
In general, a MIDI file is not defined on a time-grid; MIDI events might be defined by machine-precision digits.
The first script will take care of time-quantization by fitting every MIDI event on a time-grid the resolution of which is equal to the minimum distance between two consecutive events that are found in the input MIDI file.
The generated MIDI inherits this time-quantization.
The dimension of the unit vectors is equal to the number of different “notes” found in the input MIDI, e.g. the chromatic scale would be encoded with 12-dimensional unit vectors. Polyphony is encoded by vector addition of simultaneous events.
Similarly to LSTMetallica, the encoded input MIDI is riffled with “BAR” every 16 unit vectors for segmentation of measures. These “BAR” markers are deleted once the nerual net output is decoded to MIDI format.
1Gated Recurrent Unit
2Recurrent Neural Network
3Musical Instrument Digital Interface
4Full address or local address.