Image credits: Jason Alden/Bloomberg/Getty Images
Can AI work backwards from a text description to generate a coherent song? This is the starting point MusicLMthe AI-powered music creation tool that Google released yesterday during its I/O kick-off.
MusicLM, which has been trained on hundreds of thousands of audio hours to learn to create new music in a range of styles, is available in preview via the Google AI Test Kitchen app. I’ve been playing with it for the past day or so, as have a few of my mates.
verdict? Let’s just say MusicLM isn’t coming to musician jobs anytime soon.
Using MusicLM in the Test Kitchen is very easy. Once access is approved, you’re greeted with a text box where you can enter a description of the song – as detailed as you want – and have the system create two versions of the song. Both can be downloaded for offline listening, but Google encourages you to “like” a track to help improve the AI’s performance.
When I covered MusicLM for the first time in January, prior to its release, I wrote that the system’s songs sound like a human artist might compose them — though not necessarily musically inventive or coherent. Now I can’t say I totally stand by those words, as it seems clear that there was some serious cherry-picking with samples from earlier in the year.
Most of the songs I’ve created with MusicLM sound passable at best—and at worst like a four-year-old’s lost DAW. I mostly stuck to EDM, trying to deliver something with a distinct structure and melody (as well as gentle, ideally). But no matter how decent – even good! – Sounds like the beginning of MusicLM’s songs, there comes a moment when the songs fall apart in a very obvious and musically unpleasant way.
For example, take this template, which was created with an “EDM song in a light, cheerful, airy style, good for dancing.” It starts out promising, with a head bobbing baseline and elements of a classic Daft Punk song. But halfway through the course, it veers off course—practically another kind.
This is a piano solo from a simpler wave – “Romantic and Sentimental Piano Music”. You’ll notice that the parts sound well and good – exceptional even, at least in terms of finger work. But then it is as if the pianist has become obsessed with mania. A mix of notes later, and the song takes a drastically different direction, as if from a new sheet of music – albeit similar to the original.
I tried MusicLM’s hand at chiptunes in order to check this out, and thought the AI might be easier with songs with a basic build. We don’t dice. The result (below), while attractive in parts, ended up as random as the other samples.
On the plus side, MusicLM does, in general, a much better job than Jukebox, which is OpenAI’s attempt from several years ago to create an AI music generator. In contrast to MusicLM, songs produced by Jukebox lacked typical musical elements such as choruses that were repeated and often contained Nonsense words. Songs produced by MusicLM have fewer artifacts as well, and generally feel like a step up in terms of fidelity.
The emergence of Dance Diffusion comes several years after OpenAI, the San Francisco-based lab behind DALL-E 2, detailed its grand experiment with music generation, which it calls Jukebox. Given the genre, the artist, and a snippet of lyrics, Jukebox can create relatively cohesive music with vocals. But songs produced by Jukebox lack the larger musical structures such as the repeating choruses they often contain Nonsense words.
Moreover, the usefulness of MusicLM is somewhat limited, thanks to artificial limitations on the stimulus side. Music featuring artists or vocals wouldn’t be born, not even in style specific musicians. Try typing a prompt like “similar to Barry Manilow” and you’ll only get an error message.
The reason is likely legal. Deepfaked music stands on an ambiguous legal basis, after all, as some in the music industry argue that AI-powered music generators like MusicLM infringe music copyrights. It may not be long before there is some clarity on the matter — several lawsuits making their way through the courts likely have an impact on AI for music generation, including a case over the rights of artists whose work is used to train AI systems. without knowledge or consent. Time will tell.
“Certified food guru. Internet maven. Bacon junkie. Tv enthusiast. Avid writer. Gamer. Beeraholic.”
Chrome adds a custom user interface for Google Password Manager
Twitch revert branded content changes after backlash
AI system creates first improvements to code sorting in over a decade – Ars Technica