The generative AI revolution of 2023 continues this week, with Google releasing its AI music maker into public beta. The company announced MusicLM back in January, and it’s capable of creating 19-second musical clips from text prompts. Think of it as the ChatGPT for music.
Unfortunately, like with most new AI products, there’s a waitlist for the MusicLM preview. However, our Editor in Chief, Josh Hendrickson, has access and was kind enough to input suggestions from the Review Geek staff. Here’s the best and worst of what it made from our prompts.
Prompt: “a driving EDM track with a strong beat that’s easy to dance to. The music is steady and at a medium pace, with a strong build up to introduce a new song”
Note: MusicLM always generates two clips of music for each prompt.
This prompt did the best overall by staying true to the prompt and delivering clear sound. The excellent quality of this track may have something to do with the fact that it’s all electronic music that doesn’t have to replicate any real-world instruments (but that’s just speculation). It might be a result of the detailed prompt. Google provides instructions on what kinds of prompts deliver the best results.
Prompt: “a Gregorian chant that moves slowly and thoughtfully, keep it ambient and easy to study to”
MusicLM also seems to do well with slow, plodding music that isn’t very complicated. While the sound quality isn’t perfect (there’s a pretty pronounced skip in the first file), I can definitely see myself getting in the study zone with these files (too bad they’re only 19 seconds long). Our EIC pointed out that this one shows that it can “almost do voices.”
Prompt: “a slow romantic ballad in the style of a 40s crooner, backed by a soft piano and stringed instruments”
This prompt was the suggestion of our news editor Andrew Heinzman. Of the results, he said the second output was “pretty solid.” And he’s right. You can hear instruments and notes clearly, and it has a nice pace. If you listen closely enough, you can hear some audio artifacts of what could be interpreted as a female voice.
This output qualifies as the worst because it totally missed the mark of the prompt. I don’t know what MusicLM’s religion is, but I’ve never heard anything like this kind of music in church growing up.
My original suggestion for this prompt was “make a epic guitar solo in the style of Richie Sambora.” But I was informed that MusicLM wouldn’t make music in the style of specific artists (probably for copyright reasons). The alternate “80s hair band” prompt didn’t really deliver what I was looking for. It’s far too soft and fuzzy. Plus, this is far from “epic.”
Another prompt-expectation mismatch. Granted, I didn’t think much about what music goes over horror movie credits before I suggested this prompt, and I haven’t had time to survey my collection. But I was expecting something scary, intense, or maybe even cathartic. This output is none of those, although the sound quality isn’t bad.
Prompt: “make an instrumental song that runs the gamut of emotions from joy and bliss to fear and dread, heavy on the strings and horns”
The sound quality on these is alright, even if a little soft. Unfortunately, I felt nothing while listening to this output. Maybe the AI has a problem conjuring human emotions—and the full gamut is a pretty big ask for 19 seconds.
Review Geek staff writer Cory Gunther suggested this prompt. He wanted something “like Tool” because “they’re great.” When the results came back, he noted that “that first one honestly sounds like Tool. Similar drum style and vocals.”
MusicLM is doing a pretty decent job overall, considering how young Generative AI technology is. We didn’t expect it to be perfect. And it’s a good bet that it’s going to get much better very soon. You can join the MusicLM waitlist by visiting Google’s AI Kitchen website.