Haebichan Jung
Sep 7, 2018 · 2 min read

Super excited to have this conversation with you! Definitely pop music is melody-driven; the melody is what defines that particular song. And we aren’t in disagreement here; my comment that melody represents that “creative” bit and the harmony the “rote” bit complements your point about the tendency of composers to focus their creativity on the melodic lines.

I think where our points aren’t fully aligned is whether the harmony is conditioned on the melody or if it’s the converse. To be honest, I think both depend on each other and we should try to model our algorithms with the subtle mutual relationship in mind. Nobody composes the melody without ever considering the harmony and no one composes the harmony without thinking about the melody. It’s just the question of degree; I think that the melody is more dependent on the harmony.

This seems to be an accepted scholarly opinion; For instance, Andrew Shin et al’s 2017 LSTM model on pop music generation is built with the same assumption in mind, where their “melody generation model is conditioned on musical input features, namely chord sequence and part information”. The results are unquestionably better than the NN model I cite in my article. The latter’s model disregards the harmony almost completely when generating the melody (if it weren’t for Magenta’s Lookback features), and the results are telling. In fact, I could have built my own machine completely oppositely, where I condition the harmony on the melody. But I deeply suspect that my own generated results would sound nothing like pop partly due to the randomized generation of chord progression.

This only addresses the first part of my project. What I’m actually more interested in is the second, interrelated aspect; what defines the collection of these songs in a genre called pop? What do all these “different” pop music ultimately share? Asking this question of similarity ultimately leads us to again, the repetition in the harmony as one defining characteristic that takes the center stage. So yes, a pop song is defined by the melody. But put together, the pop genre as a whole is largely defined by the shared similarity over the harmony. That’s what my website experiments with and proves; that the shared similarity over the harmony is what allows music generated by the data to sound something like pop, not some random noise, regardless of the sample size.

Regardless of these points, I’m super excited to learn more about the music theory library you are working with. Please keep me in the loop in regards to its development. I would love to learn more about it.

    Haebichan Jung

    Written by

    Project Lead @ TDS | Data Scientist @ Recurly