Mathematical Music by Nikita Braguinski

Author:Nikita Braguinski
Language: eng
Format: epub
Publisher: Taylor and Francis
Published: 2021-01-18T00:00:00+00:00

1 Chords

Using standard settings and the numbered structure recommended by its authors, I provided the system with the following prompt:

This is a music harmony generator that creates a sequence of chord symbols.

1: C Am F G C

2: Dm Gm A7 Dm

3: G C6 D7 G

4:

The output of the system was:

F G A C F

5: Dm Am7 G6 Dm

6: G C F A C7

7: F Dm7 G#dim Dm

8: G Dm7 Gm A

9: F G6 G7 C

10: C Am G C

In my prompt, I used keywords that make it clear that I want the system to generate a very specific kind of text, the chord symbols such as C or Am which are a standard part of Western theory and notation, especially in popular music. I also used very common sequences that follow a widely known tradition of Western music theory that recommends the choice and order of chords in a given key. It can be said with certainty that my example sequences and many similar ones appear very frequently on the internet on websites and in books dedicated to music theory, and were therefore a part of the initial data on which GPT-3 was trained.

In the output, it can be seen that the system reacted correctly, providing seven additional sequences of chords, and numbering them accordingly. Only one of them directly repeats one of the initially provided examples (C Am G C). All the others follow the prompt in a more “fuzzy” way, containing four or five chord symbols. All these symbols do exist in music theory, so the system did not make up a chord like Z1 which would only superficially resemble the pattern of a letter with a number visible in the prompt. Some of the generated sequences also more or less follow the traditional harmonic movement which I employed in the prompt while others (like the number seven) can be seen as either more adventurous or more random, depending on the user’s point of view.

Given that chord sequences expressed as symbols made from letters and numbers are also a kind of text, this experiment can be seen as a success. I had a “creative” task (generating chord sequences), and I more or less accomplished it with a minimal amount of additional training.

It is true that this task can be solved much more efficiently using much simpler techniques such as randomization in the style of Athanasius Kircher’s wooden box from the 17th century (see the chapter on the early modern period). And even machine learning itself can be put to more effective use by training a model not on everything available online, but on sequences of chords and nothing else. This simpler system would also arguably deliver a much more robust and believable imitation of human harmonic thinking. Yet, all these approaches mean additional work, and, compared to a model that is less efficient but already available, it is very well possible that a user will turn to the latter.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.