Tech Gadget Central Latest Tech News and Reviews
Meta released an open-source AI-powered music generator to compete with Google.
Meta’s MusicGen tool, a demo of which can be found here, can convert a text description (e.g., “An ’80s driving pop song with heavy drums and synth pads in the background”) into about 12 seconds of audio. MusicGen can be “steered” by reference audio, like a song, to follow the description and melody.
We present MusicGen: A simple and controllable music generation model. MusicGen can be prompted by both text and melody.
We release code (MIT) and models (CC-BY NC) for open research, reproducibility, and for the music community: https://t.co/OkYjL4xDN7 pic.twitter.com/h1l4LGzYgf— Felix Kreuk (@FelixKreuk) June 9, 2023
Meta says MusicGen was trained on 20,000 hours of music, including 10,000 “high-quality” licensed tracks and 390,000 instrument-only tracks from ShutterStock and Pond5, a large stock media library. The company has made pre-trained models available to anyone with the right hardware—primarily a GPU with around 16GB of memory—but not the code it used to train the model.
MusicGen’s performance? Well—but not well enough to replace human musicians. For basic prompts like “ambient chiptune music,” its songs are melodic and comparable to Google’s MusicLM. They won’t win awards.
“Jazzy elevator music” from MusicGen:
MusicLM says:
To confuse MusicGen, I gave a more complicated prompt: “Lo-fi slow BPM electro chill with organic samples.” MusicGen produced a Lofi Girl-worthy song, surprising MusicLM.
MusicGen sample:
And MusicLM:
To change things up, I used both tools to create a George Gershwin-style piano ditty. Google blocked artist-specific prompts in the public version of MusicLM to avoid copyright issues with generative music tools.
MusicGen doesn’t filter. But “Background piano music in the style of Gershwin” results were disappointing:
Generative music is improving (Riffusion, Dance Diffusion, and OpenAI’s Jukebox). Major ethical and legal issues remain. MusicGen “learns” from existing music to create similar effects, which some artists and generative AI users find unsettling.
Homemade tracks using generative AI to recreate familiar sounds are going viral. Music labels have successfully flagged them to streaming partners, citing intellectual property concerns. However, the copyright status of “deepfake” music is unclear.
Guidance may be forthcoming. One lawsuit involving artists whose work is used to train AI systems without their consent may affect music-generating AI.
Meta, which doesn’t restrict MusicGen’s use, says all the music MusicGen was trained on was “covered by legal agreements with the right holders,” including a Shutterstock deal.
