New Meta AI Music Tool Is Trained on 10,000 Hours of ‘Licensed Music’

Meta has unveiled MusicGen, a brand new music-generating software that may flip textual content prompts into audio recordings utilizing AI.

MusicGen was educated utilizing 20,000 hours of music, in keeping with the tech large, together with 10,000 hours of “high-quality” licensed songs and 390,000 instrumental tracks.

In accordance with Meta, all of the music MusicGen was educated on was “lined by authorized agreements with the suitable holders.” However these sources are aren’t music trade giants like Common Music Group or artists like Taylor Swift. The coaching set comes from media libraries like ShutterStock and Pond5.

Meta acknowledged “the potential lack of variety within the dataset we used, which comprises a bigger proportion of western-style music.”

Utilizing prompts reminiscent of “pop dance observe with catchy melodies, tropical percussion, and upbeat rhythms, excellent for the seashore,” or “acoustic folks music to play throughout roadtrips, guitar flute choirs,” MusicGen is ready to create quick clips of music. It will also be “steered” by utilizing references to particular eras or songs.

Meta has additionally confirmed longer tracks will be created by way of MusicGen, so creators would be capable to use the preliminary music observe as a reference level. MusicGen was launched by way of Github over the weekend, and a demo is offered to attempt by way of Hugging Face.

Researchers have additionally printed a paper outlining how they educated MusicGen and the “moral challenges” that “giant scale generative fashions” create.

“Generative fashions can symbolize an unfair competitors for artists, which is an open drawback,” stated Meta’s researchers. “Open analysis can make sure that all actors have equal entry to those fashions.

“By way of the event of extra superior controls, such because the melody conditioning we launched, we hope that such fashions can turn out to be helpful each to music amateurs and professionals,” the corporate continued.

Using generative AI in music has already made headlines and sparked debate, from Grimes advocating “killing copyright,” to common artists like Oasis and The Weeknd going through AI-generated facsimiles that elevate questions on music licensing and content material possession. Musician Paul McCartney is utilizing AI to craft a never-before-released Beatles music.

Final month, Google’s “experimental AI” software MusicLM was made accessible to the general public after being unveiled in January. The software is ready to generate music from textual content prompts and even simply buzzing, with customers then offered with two completely different variations of the requested music. The concept is that customers will be capable to vote on which one they like, which is able to assist “enhance” the AI mannequin. 

In accordance with Google, MusicLM was educated on 5 million audio clips, amounting to 280,000 hours of music.

“We imagine accountable innovation doesn’t occur in isolation,” Google stated in a weblog submit earlier than confirming they’d been “working with musicians and internet hosting workshops to see how this know-how can empower the artistic course of.”