MIDI Language



Paper: Pop Music Transformer: Beat-based Modeling and Generation of Expressive Pop Piano Compositions: code

This is a modified version with an extension of multi-instrumental support.


Convert Midi into event sequence, and represented by mapped integer array.

This could send to NLP models for AI auto music composition.

Due to this project considers more about music structures as well as its chord and melody on higher level, including note, drum, tempo, musical instrument (program in midi) and its expressions (tempo and velocity), rather than digging into too much details like sound source & direction, instrumental performing techniques (such as, bend sound, piano sustain pedal, violin overtones), the language of MIDI is design this way (see chapter Details below).


See language.py, it contains procedures:

  • load w2i (word to integer) and i2w (integer to word), for not calculating it every time;
  • encode midi to iteger array, each object handle one mid file;
  • decode integer array to midi, each object handle many results and export to mid files;

The code language.py has arguments:

  • input: input file of audio file to encode/decode;
  • output: output file of audio file to encode;
  • train: if have, it will switch to training mode with variations (data augmentation);

MidiEncoder data augmentation:

  • pitch_variation_range: a random pitch shift within a range for whole midi;
  • velocity_scale_variation_range: a random note/drum velocity scale for whole midi;
  • velocity_noise_scale_variation_range: a random note/drum velocity scale for each element within midi;
  • tempo_scale_variation_range: a random tempo change for whole midi;

MidiDecoder needs numerator and denominator time signatures for reconstructing midi files.


Event Structure


  • Bar
  • Position (0~split-1)


  • note:
    • Note
    • Program (0~127)
    • Pitch (0~127)
    • Velocity (0~127)
    • Duration (0~split*bar_scale-1)
  • drum:
    • Drum
    • Program (0~127)
    • Pitch (0~127)
    • Velocity (0~127)
    • Duration (0~split*bar_scale-1)
  • chord:
    • Chord (chroma_name:chord_name)
  • tempo:
    • Tempo_Class (T0~Ti)
    • Tempo_Value (0~59)


View Github