Meta’s MusicGen Trained On 20,000 Hours Of Licensed Music, But Is It Any Good?
We’re live in a series of technology hype cycles. If you remember, there was blockchain, crypto, NFTs, Web3, and the Metaverse just in the last few years, and the world seemed to swim around them. Each of these brought the familiar “Music is going to change forever” call from the various parties involved, but (no surprise here) that never happened. We’re currently in an AI hype cycle in the music industry that promises a lot, but fails to deliver much of the time. Case in point, Meta’s (Facebook parent company) new MusicGen AI text to music generator.
Meta claims that MusicGen has been trained on 20,000 examples of licensed music, and 390,000 instrument-only tracks from ShutterStock and Pond5. This makes it different from other large language models in the fact that the training material isn’t just scraped from the Internet with no regard to licensing.
You have to hand to it Meta in that they appear to have gone out of their way to make sure that everything was on the up-and-up copyright-wise, although I think the reason was to avoid being pulled into new costly lawsuits instead of being benevolent.
Let’s Try It
But like most music AI’s on the market, the hype far exceeds the reality. MusicGen is a text to music generator, so all you have to do is describe your music (“80’s soft rock using a mandolin”) and in theory it will crank out music based on your description.
The problem is that you only get 12 seconds of music, and most of the time it’s way off the mark. I asked it simply for “Swedish Death Metal” (usually I’d make the description much more detailed) and after taking a couple of minutes to think, it came back with something so far off the mark I don’t know what to even call it. Any kind of metal it was not.
I asked it for “Dark folk music with a mandolin,” and after taking even more time to think, it generated something dronelike that wasn’t very dark, and only approximated a mandolin. I give it points for “Speed metal like Anthrax” where it approximated the machine-gun double bass drums and guitar distortion, but having to wait 2+ minutes for 12 so-so seconds of music isn’t much of an incentive to keep using it.
It might be better if you use the Drop Audio Here upload to train it on what you want, but frankly I don’t want to be complicit with training it on unlicensed material, or exposing any music that I’m connected with to an AI, especially one owned by Meta.
AI Is Not A Sign Of Quality
I’ve been intensively playing with AI music apps for quite some time now in preparation for my upcoming AI For Music Production Course, and I have to say there are many apps and plugins that prove to be a huge disappointment in what they deliver.
They either have a very steep learning curve, have a limited output resolution, have restrictions on what you can monetize or copyright, or just plain crap output, along with a host of other limitations. Then there are the few that make music production faster and easier, and can provide some ideas that you’d never thought of.
Remember, just seeing “AI” attached to an app doesn’t necessarily mean it’s any good.
Click here to try MusicGen for yourself.
