Evolving your own music for fun and no profit

Vik Paruchuri

What's evolving music?

Take existing music
Learn patterns from it
Use the knowledge to make new music
Judge your new music

Why?

Because it's cool
See above reason

Listen in

Sample tracks: 1 2 3

Midi format

[midi.ProgramChangeEvent(tick=0, channel=0, data=[0]),
midi.NoteOffEvent(tick=20, channel=9, data=[42, 64]),
midi.NoteOnEvent(tick=28, channel=9, data=[38, 90]),
midi.NoteOnEvent(tick=0, channel=9, data=[42, 90]),
midi.NoteOffEvent(tick=20, channel=9, data=[38, 64]),
midi.NoteOffEvent(tick=0, channel=9, data=[42, 64]),
midi.NoteOnEvent(tick=28, channel=9, data=[42, 70]),
midi.NoteOffEvent(tick=20, channel=9, data=[35, 64]),
midi.NoteOffEvent(tick=0, channel=9, data=[42, 64]),
midi.NoteOnEvent(tick=28, channel=9, data=[42, 58]),
midi.EndOfTrackEvent(tick=0, data=[])]

MIDI stored as byte code

midi

Python-midi to the rescue!

Parses midi into python data structures
Found at https://github.com/vishnubob/python-midi

Step 1: ACQUIRE MUSIC

Use scrapy -- http://scrapy.org/
Write crawlers to scrape sites
Download .mid files

Step 2: Render tracks

MIDI has to be "rendered"
Output to .ogg files

Step 45: Markov chain everything!

Markov chain

Construct chains

Make matrices for pitch, velocity, tick per-instrument
Rows and columns are numbers
Cells contain # of transitions from row state to column state

Note "phrases"

Music has structure
Phrases are basic units of it
I have almost no music theory experience
Let me know after if I'm wrong
Pull out sequences of 8 notes

Making a track

Select the instrument
Randomly select a starting phrase
Use markov chains to pick next phrase
Repeat until specified length

Making a song

Pick instruments that are good together
Put tracks from those instruments together
Add a tempo track

Evaluating song

Render midi to ogg file
Use fluidsynth, oggenc, and sox

Look at the audio

audio

Read audio in

$$\begin{bmatrix} 2.35185598e-05 & -1.04448336e-05\\ -3.46823663e-06 & -3.73403673e-05\\ -2.69492170e-06 & -1.44758296e-05\\ 9.47549870e-06 & 2.09419904e-05\\ -2.70856035e-05 & 3.44590421e-06\\ -3.01332675e-05 & 2.74870854e-05\\ -1.44664727e-06 & 7.49632018e-05\\ -3.80197125e-05 & 2.56412422e-05\\ -5.61815832e-05 & -1.29676855e-05\\ -4.73532873e-06 & 3.69851950e-05 \end{bmatrix}$$

Feature calculation

Calculate a bunch of features
Split song into "bins"
ZCR, slope, peak count, and many more
154 in all
Put into matrix

Scoring songs

Compare generated song with existing songs
Used random forest at one point to predict
Simpler to use cosine distance
Lower distance is better

Semi-genetic algorithm

Create candidate songs
Score them all
Pick the best ones
Remix songs by swapping tracks
Generate some new songs

Generations

Evolve multiple generations
Try to minimize distance from existing songs
Algorithm finds local minimum quickly
More than 3 generations gets duplicate "best" song

Ways to make it better

Hand-label some songs, and use them for supervised learning
Get more training data
Generate better features
Do more in the genetic algorithm than just remix

Open source

Available at https://github.com/VikParuchuri/evolve-music2
This presentation available at: https://github.com/VikParuchuri/pydataboston2015
Want to learn how to do things like this? Check out www.dataquest.io!