MusicLM by Google got it all wrong! Why?
Artificial intelligence in music - Google's new tool, is it worth it?...
On January 2023, Google released a pretty mediocre tool for creating music using artificial intelligence, and everyone is going crazy about it!
I have been receiving messages about this topic for several months now, and recently there has been an influx of information, much of it inaccurate, regarding Google's MusicLM tool. This tool is said to use AI to generate music, similar to GPT's language generation capabilities.
Therefore, I have taken it upon myself to conduct research on the subject of AI-generated music and present my organized thoughts and opinions here.
To begin with, there have been several endeavors to develop engines that utilize AI for generating music until now.
Below are four examples:
MuseNet, MusicLM, Beatoven, and AIVA.
Let's take a closer look at each one of them.
1. OpenAI, who recently stunned the world with ChatGPT, tried it with MuseNet, then abandoned it in 2019 after not really achieving what we know from a drawing or text.
2. AIVA, a company that mainly deals with music for media and music for libraries.
They approached me for advice, by the way, and I honestly told them that the music didn't sound good enough (I will elaborate on this important matter later, as most companies have missed it so far).
3. A small company with a product called Beatoven, which got its name from a combination of beat and Beethoven, in case you missed it😀.
4. Google's new engine, MusicLM, which hasn't really been released to the public yet, meaning you can't see how it creates music on its own, but you can only hear existing examples that they created in advance.
The same thing, by the way, that OpenAI did with MuseNet, they did not open it for public use but only showed a proof of concept.
The reason that Google has not released the tool for commercial use is that they currently have issues of copyright infringement with one percent of the engine's output, meaning that the engine uses melodies and chords from music that they were trained on, which is a copyright infringement, of course.
According to them, it would be very difficult to bring this down to zero percent, so it does not seem like they will release it as a commercial product anytime soon, it's still only in the research phase.
Now that we've covered the background of these initiatives, let me share my personal perspective.
All existing generative music engines, as of now, are far from what we know from drawing engines such as DALL·E 2 by OpenAI, and text engines such as ChatGPT.
And in my opinion, it will take a lot more time, maybe five to ten years, until they crack the issue, because of the following three difficulties.
These are Legal, musical, and artistic difficulties:
1. The legal difficulty - copyright infringement.
The reason Google has not released its tool to the public is, as mentioned, the concern about copyright infringement, not as written in VentureBeat, Calcalist, and other places I read. And copyright in music is one of the most complicated things in the world!
No one can really answer the question of "how many copied notes are considered a copyright violation." I once heard someone say seven notes. And someone else told me five seconds of the music.
But there is no clear answer to this; it depends on the song, its success, and the judge who is supposed to decide on it😀
Just look at the case of Doron Medalie, who is still paying Universal Music fifty percent of the royalties from the song Toy.
Because there is a similarity to the song 'Seven Nation Army';
You're wondering how many similar notes are there? Five notes! I counted.
Here is the song "Toy", which as known, won in the Eurovision song contest:
And here is the song "Seven Nation Army", feel free to judge for yourselves!
So, would you give 50% of "Toy" royalties, just because of these five notes?!
So why does Doron Medalie give them fifty percent of the profits?
He explained in a podcast that he realized that a ten-year legal battle against Universal would cost him much more, so he decided on this compromise, even without going to court.
There is also the example of Ed Sheeran and the song Shape of You. After a long and expensive legal battle, Ed Sheeran ultimately won the case, but it earned him millions.
Here's the song Shape of You, which already has 6 billion views (!) as of January 2023:
The point here is simple:
If Google's engine were to produce, for example, 1000 songs every week (or day) in which there is copyright infringement, lawsuits would start here that would never end. And Google doesn't want to get into that. That's the legal issue.
2. The music itself - in my opinion, the music produced by all these engines is simply not interesting enough.
Music is different from a painting in that it exists on the axis of time. The excitement of music comes from the fact that it has a certain peak, which in many cases comes in the last third of the song or musical piece.
And this aspect, of a story in sounds on the axis of time, all these engines have missed in my opinion.
Maybe if Google paid me a million dollars, I would share with them what we teach at our Jerusalem Academy of Music, about how to create interesting, persuasive, and storytelling music. In fact, a million dollars is too cheap for it😀
But until then, these engines create simple loops that repeat, which do not sound good from a musical production standpoint (and we'll talk about that in the next section), and also do not create a convincing and emotional story!
3. Musical production
When we listen to music today, we primarily hear the musical production, the sound quality, the sample quality, and how authentic and human they sound.
What Google and OpenAI do not yet know is that for a computerized violin to sound good, my composer friends and I work for hours on the computer and on the music production software, designing every sound and every millisecond within its sound so that it does not sound "computerized," but rather as human as possible. And that is something that cannot be quantified, it is something that needs to be heard with the ear.
Everything I've heard so far from these engines, from a musical production standpoint, sounds like the simple organs of Yamaha from the 1980s, that is, very simple sounds, especially when it comes to instruments with long sustain, such as violin, strings, wind instruments, and piano (percussion instruments are easier to do, and they actually sound good even in these engines).
So there is still a long way to go until artificial intelligence truly succeeds in creating music. It's probably a tougher nut to crack than other things. Will they solve this in the coming years?
Perhaps because I'm a musician, I expect more. I assume that by 2030 we will already be hearing generative music that sounds good, tells a story, and somehow manages not to violate copyright.
But until then, my dear friends, keep making music!! 🎵
That's me in the studio. Not feeling (yet) threatened by AI that will replace me😀
In conclusion, AI-generated music is an exciting new frontier in the world of music production. While the technology is still in its early stages, it has the potential to revolutionize the way we create, listen to, and experience music. As the technology continues to develop and improve, it will be interesting to see how it transforms the music industry and shapes the future of music. Whether you love it or loathe it, there's no denying that AI-generated music is here to stay, and it's an exciting time to be a part of the ever-evolving world of music.
Thanks for reading!
Is this your first time on this blog? It's nice to meet you!
Let me introduce myself -
I'm Dr. Amit Weiner, a composer and pianist. I'm also the head of the Department of Multidisciplinary Creation at the Jerusalem Academy of Music and Dance, which is one of the top ten departments in the world for multidisciplinary composition.
I work with some awesome companies like Universal, Sony, Warner, Megatrax, and more. Let me know if you have any questions or just want to say hi!