Meta introduces text-to-music AI generator, ‘MusicGen’

Over the past few months, AI tech has been all the craze with surveys showing how 73% of producers think AI music generators may replace them, over 60% of musicians are using AI to make music and how AI replicated the voices of Drake and the Weeknd.

Adding to the increasing number of news about AI advancements, Meta has introduced a new AI generator ‘MusicGen’ that’s able to convert text to melody.

‘MusicGen’ features a straightforward interface, allowing users to type in a prompt describing the type of music they would like generated. It also includes the option to attach an already-existing track that can be modified into whatever your prompt states — such as plugging in the Star Wars theme song and turning it into a Peggy Gou-style banger.

The AI generator uses an open-source deep learning language that Meta has trained by using over 20,000 hours of music, which includes 10,000 hours of licensed tracks and 390,000 instrumental tracks. The final samples can range from 10-30 seconds and most definitely lives on the more basic scales of harmonies.

This is what we got for the prompt “90s jungle techno rave track with an electro melody of 140bpm”:

Read this next: Kito and Grimes share new AI collaborative track made using "Elf.Tech"

Generating the sample took quite a bit of time; around 30 minutes for a 15-second sample. Also, it took up a lot of GPU power and for those with older computers, it would probably take a lot longer.

Read this next: Universal asks streaming platforms to block AI access of music catalogue

Result times are also dependent on how many people are generating samples at once — it was a lot faster to generate these at night (in Asia) compared to the daytime.

There are two ways of using this interface; one is to jump onto the site on Hugging Face’s API. The other option, for the more tech-savvy, is to get a code and download it onto your PC.

Miki Kitasako is Mixmag Asia’s Social Media and Content Producer, follow her on LinkedIn.