Assignment 5: Music with Markov Chains

Objectives

Learn how to generate polyphonic music with Markov Chains.

Description

A Markov Chain is a statistical model of a system that changes stochastically from one state to another. It can be applied to generate random sequences  that follow a probability distribution learned from a dataset. In this assignment, you will use a Markov Chain to generate polyphonic music.

For this assignment, you will use a base code to load and parse MIDI files to create your own dataset. To load a list of midi files using this base code, you have to create a .json file that defines a list of MIDI filepaths as follows:

data/midi_files.json:
{"file0": "data/midis/Super_Mario_Bros_Underworld.mid",
 "file1": "data/midis/Super_Mario_Bros_MainTheme.mid",
 "file2": "data/midis/Super_Mario_Bros_Underwater.mid"}

The base code provides a python script that can generate this list given a directory with MIDI files in it:

python3 midiPlayer/py/load_midi.py --dat data/midis/ --out data/midi_files.json 

Once your create your .json file, you can load all the MIDI files describe inside it using the following code:

let midiPlayer = new MidiPlayer();
midiPlayer.loadMidis("data/midi_files.json", onMIDIsLoaded);

Where the first argument is the path of your .json file and the second is a callback function that is called when all the MIDI files have been loaded. This function passes the list of MIDI files as piano rolls as argument so you can parse them, for example:

function onMIDIsLoaded(pianoRolls) {
  let pianoRoll = pianoRolls[0];
  let midiText = midiPlayer.pianoRoll2Text(pianoRoll);
}

In this example, "midiText" contains a string representation of the first MIDI file in your .json file. In this assignment, a MIDI file is encoded as a string where each time step of MIDI's piano roll is separated by the character "." and every note is represented by a symbol of the form [pitch]_[duration], where pitch is a integer from 0 to 127 and duration can be any integer. In this encoding scheme, space is a separator of symbols. For example, consider the piano roll below:

The first 20 time steps of this piano roll are encoded by the following string:

62_4 . . . . . 61_4 64_4 . . . . . 60_4 66_4 . . . . . 59_4 67_4 . . . . .

The provided base code can also be used to play and visualize your generated music pieces as shown in the example above.  Once you have a generated piece as a string  you can play and visualize it as follows:

let midi = midiPlayer.text2Midi(midiText);

let midiData = midiPlayer.parseMidi(midi);
let pianoRoll = midiPlayer.notes2PianoRoll(midiData.duration, midiData.notes);

midiPlayer.setPianoRoll(pianoRoll, tsCallback);

To accomplish this assignment:

1. Clone the base code  from your class GitHub.
2. Download from the internet a large number (e.g. 500) of MIDI files of piano music from the same music style (e.g. classical music, jazz,  video game soundtracks, etc).
3. Create a .json file with all these download MIDIs using the load_midi.py script.
4. Load (in sketch.js) all these files and parse them as strings.
5. Implement (in "markov.js") a Markov Chain class which supports training a markov chain on a given dataset and sampling from it to generate new pieces on the style of the dataset.
6. Use (in "sketch.js")  your Markov Chain to generate a new piece every time one reloads your sketch.  User should be able to play this piece and visualized it as a piano roll on the canvas.
7. Using your Markov Chain class, generate 5 different pieces using different amounts of data (e.g. 100, 200, 300, 400, 500). 

Resources

• Classical Piano MIDI files
• The MAESTRO dataset
• Video Game Music MIDI

Submission

Host your source code on Github and upload the following on Canvas:

1. A zip file with the final version of your code.
2. A zip file with the 5 different music pieces generated using different amounts of data.
3. A link (as a comment to your submission) to this assignment's github page. This page should have examples (.png or .gif) of outputs of your Markov Chain with a brief explanation (1 paragraph) of  how this Markov Chain generates music.