ChopAIn: His Nocturnes According to AI

This article is related to the forthcoming publication: 

Encounters between music, education and technology (Murillo, Tejada, Marín, Riaño, González, Añó, & Arnal, in press), 

part of the ARTSLAB Contemporary Artistic Education series, published by Publicacions de la Universitat de València. 

Once the book is published, this entry will be updated with the final details and a link to access the publication.

Introduction

In this post, I decided to test and compare different generative audio AIs with some very simple tasks.

I asked them to:

Create a nocturne in the style of Chopin
Create a Chopin nocturne played by a progressive rock band
Create a Chopin nocturne in the style of Bach
Create a Chopin nocturne in the style of a classical Indian raga
Create a Chopin nocturne as if it were an opera

For all the AIs, I used the same prompt (in English) — the simple phrases above — to test their ability to recognize a composer’s style. I deliberately avoided crafting a more complex prompt where I would describe the characteristics of the composer myself. Some of these AIs also offered advanced options, but I chose not to use them, as the goal was to compare them “out of the box.”

The objectives were, on the one hand, to see if the AIs could recognize what a Chopin nocturne is and create an audio file where the characteristics of these pieces were recognizable, and on the other hand, to explore the interactions between the understanding of a Chopin nocturne and other styles.

For example, the use of a progressive rock band (a genre known for many “covers” of classical music) was to see how it could merge both styles.

Bach was one of Chopin’s influences, so I wanted to observe how the AI would “blend” Chopin’s style with that of an older historical period—essentially, an anachronistic experiment.

With the Indian classical music raga, the aim was to see how the AI mixed styles from different cultures.

And with the opera, the idea was to analyze how it handled transforming an intimate pianistic genre into a grand orchestral one (with the hope of seeing how Chopin’s lyrical piano writing would be translated into sung voices).

The AIs that were compared are:

LimeWire
MusicGEN (Meta)
MusicFX (Google)
StableAudio (Stability)
Suno
Udio

As you can see, I prioritized models from major companies (Meta, Google, and Stability) and then included LimeWire (due to its historical background, which could prove interesting) as well as Suno and Udio (as examples of commercial models that have been emerging recently). There are some other commercial examples (such as Beatoven and Aiva), but these generally generate results quite similar to those produced by Suno.

Chopin Nocturne (in the style of Chopin)

In this first series, we come face to face with what each AI understands about what Chopin is.

MusicGEN

Since I wasn’t sure about the limitations of the version and didn’t want to run out of “credits” for free creations, on the first try I generated a short one (20 seconds), and then I used other prompts to request longer versions (50 seconds). For that reason, there are two examples for each in MusicGEN: one “short” and one “long.”

Short

Not bad at all:

the piano is recognized as the main instrument.

Although the melody seems unsure of where to go, the cadential “ornaments” and “chromaticism” typical of Chopin do appear.

Long

Once again, Chopin is quite recognizable due to the same characteristics already described. We have more dotted rhythms, and the accompaniments are also relatively Chopin-like. In the middle, there is a rather “dizzy” moment without direction, and it seems like there is some orchestral sound in the background.

LimeWire

Right from the start, the type of chosen timbres is quite confusing. Moreover, Chopin is nowhere to be found (except maybe at 0:08).

Suno

Suno actually creates two versions for the same prompt, which is why there are two examples for each case in Suno.

Version 1

Seriously? Einaudi? Yiruma? Chopin is nowhere to be found.

It definitely has a commercial finish on several levels:

This certainly sounds more like a real piece compared to the previous examples.
There is a more or less present structure with its climax (around the 3-minute mark).
The music it generates is more “commercial” in style than what I asked for: Chopin.

Versi0n 2

We’re still in the same place… Neo-classical new-age. There’s even a sound entry (just before the first minute) of… a mini string orchestra? A synthesizer?

Stable Audio

The sound is still “tinny,” but there’s more Chopin here than in the previous Suno examples. The melody is accompanied, but the harmonic progressions aren’t in his style. There’s more chromaticism in the harmony than in the melody, although there are glimpses of attempts at a Chopin-like melody.

MusicFX

With MusicFX, two versions are also created by default.

Version 1

Maybe if I had asked for Satie, it would have fooled me.

Version 2

Here there’s “a bit more of Chopin,” although it sounds like the soundtrack to a romantic movie from the first half of the last century. But it still has the same timbre as before, with plucked strings.

Udio

Udio also creates two versions by default.

Version 1

Here we do get a sense of Chopin, although somewhat unfocused, as if he were improvising rather than playing a clearly defined piece. However, the atmosphere, accompaniment, progressions, and melody; practically everything is Chopin.

Version 2

Surprise! The right hand is a cello! Though perhaps what impresses me most is how the model has internalized the (physical) breaths of the cellist as part of the music, and you can hear them at different moments. And while it breathes more often than would normally be expected, it does so in places where it would naturally occur (at the end of some phrases or just before the start of new impulses).

Chopin Nocturne played by a progressive rock band

Now that we know what each AI understands as Chopin, let’s see how successful (or unsuccessful) they are in transforming Chopin into a cover by a progressive rock band.