If you’ve played around with ChatGPT at all, you know how impressive (and terrifying) AI can be. ChatGPT can generate just about anything you ask it to, with prompts as simple or as complex as you can think of. Well, Google has figured out how to do the same. But rather than spit out poems or computer programs, Google’s AI writes music, and you can preview its creations right now.
Google’s AI, dubbed MusicLM, is designed to produce multiple minutes of high-quality music based on text prompts. While AI music isn’t necessarily new, the company claims its program produces tracks that are more accurate to the initial text prompt and are better quality than previous models. Big claims, of course. But, based on these previews, they might just live up to them.
The first batch demonstrates how MusicLM can generate music from rich captions, similar to how OpenAI’s ChatGPT generates its famous responses from user requests. Google can feed MusicLM a caption, such as, “The main soundtrack of an arcade game. It is fast-paced and upbeat, with a catchy electric guitar riff. The music is repetitive and easy to remember, but with unexpected sounds, like cymbal crashes or drum rolls,” and the AI will process a track that meets those criteria.
My personal favorite from this section is the song generated from “Funky piece with a strong, danceable beat and a prominent bassline. A catchy melody from a keyboard adds a layer of richness and complexity to the song” on the second page, which sounds like something out of Stardew Valley.
Where uncanny valley really sets in, though, is whenever Google asks MusicLM to produce vocals. Many of Google’s examples here have AI vocals, and while some of them sound “computerish” for lack of a better word, others sound a little too close to home. When you listen with the full knowledge that the voice is 100% artificial, it’s a bit unsettling. Google even asked MusicLM to generate a rap track, and, to be honest, the “rapping” is eerily realistic, even if none of the words they’re rapping are real.
If you really want to be creeped out, scroll down to the “Text and Melody Conditioning.” Here, you can hear a series of voices and instruments doing their best with different melodies, and the results are wild. The very first option sees MusicLM humming “Bella Ciao,” and it sounds way too good. Side note: The “Tribal drums and flute” version is giving White Lotus.
We also have “Long Generation,” which produces five-minute long tracks from short descriptions, like “melodic techno” or “relaxing jazz.” But things get really interesting with “Story Mode.” Google feeds MusicML multiple captions, and the AI adjusts the song based on the given caption. It starts with 15 seconds based on “time to meditate,” before the caption switches to “time to wake up.” The music smoothly changes in tune, as if it was starting up a new verse, with some creepy vocals to boot.
Creepy vocals aside, it’s fascinating to see the AI alter the song without having to change the song. For the most part, it incorporates the new caption into the song, without feeling like you’re jumping into a new tune altogether. It’s particularly impressive with the third example, which sees the AI jump from “pop song” to “rock song” to “death metal song.” While “rock song” doesn’t particularly sound like rock, “death metal song” is simply fantastic. I imagine AI Death will be a big hit with metal heads.
Another interesting experiment Google did was feed MusicLM descriptions of famous paintings to see what music it would produce. If you ever wanted to know what Vincent Van Gogh’s “The Starry Night” would sound like as a song, well, now’s your chance. It’s also fascinating to click through how MusicLM generates specific instrument and genre sounds. The electric guitar is on point, as is “British Indie Rock,” complete with a dramatic turn at the end of the preview.
There’s more to explore from this limited preview of MusicLM, too, from “Accordion Solos” to comparing different results from the same prompt. Google also published a 15-page research piece on MusicLM, if you’re interested in reading the technical details of the system. While I hope the company releases a beta for the public to try, browsing through these samples is the next best thing.