Computer written songs?

The “magic chord” feature had me thinking. It analyzes not only the chords that come before it and after it, but also the melody, to figure out which chords would sound good. That made me wonder, what would happen if there was also a similar feature for melodies, to predict the next likely note?

Then that made me think, if you combined all of those, wouldn’t the program be writing its own songs? It reminded me of some other programs people have made to do something similar. I remember one person made one that analyzed classical compositions of certain composers and created music in the style of those composers.

With all of this data, surely we can create our own program that writes its own songs. In fact, certain artists on here, such as The Beatles, have so many songs that the program could probably just make its own songs in the of that one artist. It would be interesting to see the results!

My guess is the staff clearly doesn’t have and won’t acquire the storage and bandwidth required to contain all melodic information from tens of thousands of Theorytabs even though at the same time the database is not large enough for Magic Chord to handle every non-supermodal chord.

Fair enough, but it would be cool if somebody downloaded all the data and made their own program outside the website that wrote its own songs.

cgMusic is a procedural music generation program, if that’s at all what you’re looking for

1 Like


We can certainly take statistics on melody notes in ThoeryTabs.

However, writing songs is probably more complicated than just stringing together chords and notes. Such an approach may ultimately produce a pleasant-sounding jingle, but ignores all of the more nuanced elements (rhythm, phrasing, etc) that make up a good song. There are certainly people that have tried to do this, but as far as I know, nobody is getting famous off of it.

That being said, if you have any specific ideas for how you would like to use TheoryTab data, we’d be happy to hear them.

Good discussion here. Melody is a lot more complicated in general. It’s not just as simple as predicting what note should come next. The rhythms syncopations, and underlying harmony are a huge part of a what makes a melody good. Melodies also have a lot of structure to them that would be hard to capture. In some sense this is true of the harmony as well. Magic chord is great at giving you suggestions for a missing chord, but if you just kept asking what to do next over and over again, you would end up with chords that work well with their neighbors, but a “song” that lacks a coherent structure. There wouldn’t be a start or an end, for example. I think this is an issue with a lot of procedurally generated music.

That said, we have considered adding some form of melodic data to trends. There’s definitely things that I think would be interesting to pursue that would add a lot of value. As @HertzDevil mentioned, some ideas might be too resource intensive to be practical, but there are certainly other ways of manipulating the data that wouldn’t be.