Data-driven Chord Progressions

Whenever I listen to music I like trying to identify chord progressions and find common (or uncommon) patterns and changes. I recently came across the McGill Billboard dataset, which is a set of Billboard Top 40 songs from the 60s, 70s, and 80s annotated with structure and chord information. This led me to build a little widget for creating 4-chord progressions.

Chord Bigrams

The chord picker is based on chord bigrams, which are just pairs of chords that appear one after the other in a song. For example, if I have this 4-chord progression:

1maj -> 5maj -> 6min -> 4maj

Then I can extract the following three chord bigrams:

• 1maj -> 5maj
• 5maj -> 6min
• 6min -> 4maj

If this progression repeated there would also be the 4maj -> 1maj bigram.

Counting Bigrams

In the widget, chords are sorted by the frequency of the chord bigram with that chord and the previous one. So if I chose 1maj as my first chord, the list of possibilities for chord 2 would be sorted based on how often a chord is observed to come after 1maj in the dataset. (This is true for all but the first chord, where a small set of "seed" chords are listed.)

This means if you're picking chords at the top of the list, your progression will sound more conventional. If you're picking chords near the bottom of the list your progression could sound more distinctive, but it could also sound more disjointed.

If you don't have a great ear for chord progressions I suggest first plugging the progression above (which appears in quite a few songs) into the widget below, hearing how it sounds, and then trying to modify it.

You can check out the completed code here.