To not be outdone by Google, Meta has launched its personal AI-powered music generator — and, not like Google, open-sourced it.
Referred to as MusicGen, Meta’s music-generating device, a demo of which might be discovered here, can flip a textual content description (e.g. “An ’80s driving pop tune with heavy drums and synth pads within the background”) into about 12 seconds of audio, give or take. MusicGen can optionally be “steered” with reference audio, like an present tune, by which case it’ll attempt to comply with each the outline and melody.
Meta says that MusicGen was skilled on 20,000 hours of music, together with 10,000 “high-quality” licensed music tracks and 390,000 instrument-only tracks from ShutterStock and Pond5, a big inventory media library. The corporate hasn’t offered the code it used to coach the mannequin, but it surely has made out there pre-trained fashions that anybody with the correct {hardware} — mainly a GPU with round 16GB of reminiscence — can run.
So how does MusicGen carry out? Nicely, I’d say — although actually not nicely sufficient to place human musicians out of a job. Its songs are fairly melodic, at the very least for fundamental prompts like “ambient chiptunes music,” and — to my ears — on par (if not barely higher) with the outcomes from Google’s AI music generator, MusicLM. However they received’t win any awards.
Right here’s the output from MusicGen for “jazzy elevator music”:
And right here’s MusicLM’s take:
Subsequent, I gave a extra difficult immediate to try to throw MusicGen for a loop: “Lo-fi gradual BPM electro chill with natural samples.” MusicGen surprisingly outshined MusicLM when it comes to musical coherence, producing one thing that’d simply discover a dwelling on Lofi Girl.
Right here’s MusicGen’s pattern:
And right here’s MusicLM’s:
To change issues up a bit, I attempted utilizing each instruments to generate a piano ditty within the fashion of George Gershwin. I say “tried” as a result of, in an effort to forestall the copyright points round generative music instruments, Google applied a filter within the public model of MusicLM that blocks prompts mentioning particular artists.
MusicGen has no such filter. However the outcomes for “Background piano music within the fashion of Gershwin,” left one thing to be desired, I need to say:
Generative music is enhancing, clearly (see Riffusion, Dance Diffusion and OpenAI’s Jukebox). However main moral and authorized points have but to be ironed out. AI like MusicGen “learns” from present music to supply comparable results, a reality with which not all artists — or generative AI customers — are snug.
More and more, homemade tracks that use generative AI to conjure acquainted sounds that may be handed off as genuine, or at the very least shut sufficient, have been going viral. Music labels have been fast to flag them to streaming companions, citing mental property considerations — they usually’ve generally been victorious. However there’s nonetheless a scarcity of readability on whether or not “deepfake” music violates the copyright of artists, labels and different rights holders.
It won’t be lengthy earlier than there’s steering on the matter. A number of lawsuits making their means by the courts will doubtless have a bearing on music-generating AI, together with one pertaining to the rights of artists whose work is used to coach AI programs with out their information or consent.
For its half, Meta, which isn’t imposing restrictions on how MusicGen can be utilized, says that each one the music MusicGen was skilled on was “lined by authorized agreements with the correct holders,” together with a cope with Shutterstock.