LISTEN: WE TRIED GOOGLE’S AI TEXT-TO-MUSIC GENERATOR—AND IT’S FREAKISHLY GOOD (EDM.COM)
MusicLM is by far today’s most powerful and intuitive AI music generator. But what are its implications?
Google is rewriting the symphony of creativity with MusicLM, its groundbreaking AI text-to-music model.
According to an academic paper published in January, the tech giant said there were “no immediate plans” to share the experimental AI tool, which lets users type descriptive prompts to generate unique songs. But they reversed course last week, when they announced in a blog post that MusicLM’s waitlist was open.
I’ve tested a myriad of AI music programs and haven’t found a single one with genuine promise, so I had my suspicions. But after just a few minutes playing around with MusicLM, any scintilla of skepticism eroded. It’s by far today’s most intuitive and dynamic AI music generator.
After opening MusicLM, you’re met with a modal containing instructions on how to compose the most effective prompts, like mentioning the “vibe, mood or emotion” you want to produce. Most notable, however, is this line: “Certain queries that mention specific artists or include vocals will not be generated.”
That seems like a veiled response to the recent proliferation of scam AI songs, the most explosive of which have used AI deepfake tech to replicate the voices of Drake and Frank Ocean, among other major contemporary artists.
Most regard copyright issues as the touchstone of the nexus of AI and music, but at the moment there’s no silver bullet to the issue—and the reckoning is only just beginning as lawsuits pile up by the day. However, it seems Google is acutely aware of the implications.
Clicking out of the pop-up takes you to MusicLM’s user-friendly dashboard, where you’re able to enter a prompt and feed it through the AI.
The tool then quickly spits out two downloadable 19-second audio clips, which are grainy and muddled—but that’s to be expected. If Google were to integrate AI-powered mixing and mastering services, it would be game over.
I started out with a simple prompt, “Deep, grimy dubstep at 140 beats per minute.” The program, which Google says was trained on a dataset of roughly 280,000 hours of music, didn’t disappoint.
In terms of the relationship between a prompt and its resulting audio, MusicLM is much more effective when writing vividly with descriptive words. William Faulkner told us to murder our darlings, but he wasn’t a musician.
Here’s the mind-blowing result of my prompt, “Ethereal, lush lo-fi with a melody evoking a warm summer day sipping rosé.” You could say I’m ready for summer.
I got another surprisingly intricate, phantasmic song after entering the prompt, “Soft, fluttery house music I would hear in a blissful dream.”