Everything you could want to know about Music and its Complexity, by using AI

Is pop as complex as jazz? What is the most - and least - sophisticated piece of music? In this article, I shall answer these pressing questions about music that music critics have been squabbling since its advent using MIDI and AI.

First though, how can AI be applied to figure all of this out? Its fairly simple. Language models like ChatGPT learn by imitation, and a byproduct of this is that they also learn to generate probabilities for what comes next too. Music is a language too (as cheesy as it might be to say). It has general guidelines, and more specific rules as you enter some genres: like dialects. We can convert music into something readable to a language model by using a file format: MIDI, which is digital sheet music. Furthermore to narrow the music down, we will only be focusing on the notes of the music: excluding rhythm, velocity, etc. This is since if a recording isn't quantized, the probabilities for the rhythm would be extremely off. Moreover, each song is split into 5000 character chunks (~500 notes), so that way if there is a 256 TB MIDI somewhere in the data, or really any big piece, it doesn't have a significant advantage over if the piece was tiny. So, with these definitions in place and a language model capable of decent prediction for music, I ran the probabilities on 20,000 MIDI files.

The most complex songs/pieces

The way this works, is that the logprobs are the accumulated probability of each token. A song with an accumulative logprob value of -8000 is 2x more 'random' on a note-to-note basis than a logprob value of -4000. Each song is split into 5000 character chunks, about 500 notes each, indicated by the leading number at the end of the track.

The most basic pieces

Overall, there are some trends with certain genres being placed in certain places. For the most complex pieces, it is actually a pretty fair share of the top 100 being a good mix of pop, classical, and jazz. On the other hand, the same could not be said about the most basic pieces. The most deterministic pieces are all rock, pop-rock, and metal. Looking through the data, I think I know 3 reasons for this.

• 1. Power-chords: By using 2 notes that will always be a 5th apart, it eliminates the probability of there being extensions to the chord
• 2. Repetition of notes in backing tracks: In addition to power chords, the songs tend to strum eighth notes, multiplying the problem of the chords already being predicable.
• 3. Songs generally being generic: Many of the songs are actually just plain, corporate made pop-rock (Though that's not to say that all are of course)

Taking a peek at some metal bands, it seems to hold true that they are affected to only be below average:

It somewhat holds true for rock bands, but not as much

Doing some more research, tracks with drums in it also have a disadvantage. When running the prediction of the model, I also kept track of the probabilities on a note-to-note and track basis. So here is a graph of the average track complexity for each note index (The green track is the drum track):

As you can see, the logprobs for each drum track in the middle, are about half of the rest of the tracks, meaning that on average, the drum track of a piece is 2x as deterministic that an actual note track. Unrelated but very weirdly though too, there are 2 massive drop offs at the note index 400. I don't know why this happens? It could be from the fact that the model may not have been trained to handle that high of context, or that maybe on average a song changes at that point. Also, as a test I evaluated this on the songs of Metallica as an example too and found that the drum track probabilities are about the same as for the rest of the tracks (Track 9, or the light teal track is the drum track) (Also note that these are the smoothed moving average probabilities of the tracks, with a 20 window size):

And The Beatles for comparison:

Which artist has the most, and least complex music on average?

For this, the data of artists was shrunken to the artists that have 10 or more unique MIDI files. Here are the artists that on average have the most complex pieces/songs:

Most deterministic artists:

Generally, same trends hold true from the song complexity tables.

Notes and Conclusion

It should be noted that this article's results should be taken with a grain of salt; especially at the extremes for places, it may not be accurate. Furthermore, the results were generated by a relatively small MIDI generation AI model, so it could change if a larger model was used. Furthermore, just because a genre or artist that you like is closer towards the bottom, it doesn't mean that their lyrics are not deep or whatever. The same goes for what is at the top. This article is flawed and more research should be done to provide more conclusive results, but I hope this serves as a decent start for more research to come.

← all research