Meta’s MusicGen AI Works with Language and Song Prompts

Meta Platforms has debuted what’s being called “ChatGPT for audio.” MusicGen is an AI music generator that can create tunes from natural language or song snippets. The company says MusicGen was trained on 20,000 hours of music, including 10,000 hours of “high-quality” licensed songs and 390,000 instrumental tracks. Meta released MusicGen on GitHub this past weekend, and is currently demoing the app on Facebook’s Hugging Face page. Visitors can generate tunes by describing the sound they want. Among Meta’s prompts: “80s driving pop song with heavy drums and synth pads in the background.”

“Not to be outdone by Google, Meta has released its own AI-powered music generator — and, unlike Google, open-sourced it,” writes TechCrunch, noting that while Meta hasn’t released the code used to train the model “it has made available pre-trained models that anyone with the right hardware — chiefly a GPU with around 16GB of memory — can run.”

Engadget says it’s “like ChatGPT for audio” and works by “letting you describe the style of music you want, drop in an existing tune (optionally) and then clicking ‘Generate.’”

Generative music has come a long way, as demonstrated by apps like OpenAI’s Jukebox, as well as Dance Diffusion and Riffusion, but legal and ethical questions swirl. At issue is the fact that music generators train on existing music and produce copycat results.

Meta says that all of MusicGen’s foundation training was done using music covered by licensing agreements, including with Shutterstock and Pond5. TechCrunch says Meta “isn’t imposing restrictions on how MusicGen can be used.”

Increasingly, AI-generated music is being put to imaginative use. “More than 50 years after the Fab Four’s final studio album, Paul McCartney says he has used artificial intelligence to create what he called ‘the last Beatles record,’” The Wall Street Journal reports.

McCartney tapped the technological expertise of director Peter Jackson to harness AI. Although McCartney didn’t reveal the name of the song, he confirmed it has John Lennon’s voice, which Jackson “was able to extricate … from a ropy little bit of cassette” on which Lennon also played piano.

“Increasingly, homemade tracks that use generative AI to conjure familiar sounds that can be passed off as authentic, or at least close enough, have been going viral,” explains TechCrunch, adding “music labels have been quick to flag them to streaming partners, citing intellectual property concerns — and they’ve generally been victorious.”

“Last month, Google’s ‘experimental AI’ tool MusicLM was made available to the public after being unveiled in January,” writes Decrypt, explaining that “the tool is able to generate music from text prompts or even just humming, with users then presented with two different versions of the requested song.

No Comments Yet

You can be the first to comment!

Leave a comment

You must be logged in to post a comment.