The hills come to life with the sound of AI-generated music as Meta launches a new language learning model (LLM): the aptly named MusicGen.
Developed by Audiocraft’s in-house team, MusicGen is like a musical version ChatGPT. You enter a short text description of the type of music you want to listen to, click Generate, and the AI quickly creates a 12-second track according to your instructions. For example, someone might say MusicGen generate “lofi slow BPM electro chill”. [song] with organic samples” and indeed, the sound sounds like something you hear on Lofi Girl radio on YouTube.
It is possible to “steer” MusicGen by uploading your own song to give the AI a better sense of structure. One of the LLM developers, Felix Kreuk, has posted some samples of what it is sounds like his Twitter profile. MusicGen can take the famous Sebastian Bach as an example Toccata and fugue in D minor then add some drum and synth beats straight out of the 80s to create a more upbeat version of the song.
Introducing MusicGen: a simple and controllable music generation model. MusicGen can be prompted with both text and melody. We share code (MIT) and models (CC-BY NC) for open research, playback and the music community: pic.twitter.com/h1l4LGzYgfJune 9, 2023
Availability
MusicGen is currently available to the public at Meta’s Hugging Face website so everyone can try. Note that unlike Google’s own MusicLM AI music generator, the Meta model cannot perform vocals, only instruments. This is probably the best solution, because MusicLM’s vocals sound very similar Seemingly. Nobody will be able to understand one thing.
For musicians, you don’t have to worry about losing your career. The AI is decent creating simple, short melodiesbut little else. In our opinion, quality is not on the same level as something made by human ingenuity. Some tracks can be quite repetitive as MusicGen goes through the same sequences over and over again. This tool can be useful for creating plain background sound for videos or presentations, but it’s nothing really engaging. The next pop hit won’t be generated by AI at least not yet.
act fast
If you’re interested in trying out MusicGen, we recommend that you act fast. First, the Hugging Face site is unstable. We had tons of AI-generated songs ready to be shared. However, the website went down while we were working on this piece, cutting off our connection to the tracks. We suspect that dead links were caused by sudden high user traffic. Hopefully by the time you read this, Hugging Face will be working properly.
The second reason is more contentious. On official GitHub pageMeta says her team used 10,000 “high quality [licensed] music tracks” and free tracks from Shutterstock and Pond5. Ever since the generative AI craze began earlier this year, artists began suing developers and platforms similarly on the ‘illegal use of copyrighted works’. The meta may soon be in the crosshairs of some crossover musicians. Even if he doesn’t get sued for using licensed music for LLM training, labels aren’t afraid of it flex your industry muscles to close this type of content.
If you’re looking for in-depth information on using AI to generate images, be sure to check out TechRadar list of best AI graphics generators for 2023.