After graphic designers, lawyers, or even journalists, it is now the turn of duty musicians update their CV urgently.
In any case, if the development of MusicGen were to continue, an entire industry could have to reinvent itself. But the first public version of this artificial intelligence, if it is slightly more advanced than its MusicLM sibling, should not immediately take the lead in the top 50 for all that.
What is MusicGen?
Meta today announced the release of a first version of MusicGen, an open source AI capable of generating music. From the user’s point of view, its operation is extremely simple: just like DALL-E or the inevitable ChatGPT, it suffices to describe to it in one or two sentences the style of music and other constraints that one wishes to see appear, and MusicGen , after two or three minutes of waiting, creates a piece that respects its codes.
As much to say it right away, there is a certain margin of progress: the pieces generated do not last for the moment more than 12 seconds, and if they more or less respect the instructions in terms of musical style, it does not generate of words and there is not yet enough to cry genius, these 12 seconds sometimes not preventing it from closing.
This is not the first artificial intelligence of its kind: last month, Google released an AI with similar characteristics, called MusicLM. And if MusicGen still has progress to make, it is still more advanced than its peers.
A risk for musicians, really?
No need, for the moment, to pass the albums that come out in an AI detector: the latter cannot yet compete. But MusicGen is only at the beginning of its existence and the recent past has shown us that the progress of AI wasted no time, and was so rapid that it terrorized its own creators.
This model was trained using over 20,000 hours of licensed music, some owned directly by Facebook, others by Shutterstock and Pond5, an online music bank. It is easy to imagine that a company like Meta has the means to considerably increase the database to train future versions of the model, while refining the technology. Moreover, the model being open source, user feedback will also allow it to sprint towards a credible AI as a music composer.
In the meantime, as it is unlikely that Meta will keep it free, do not hesitate to have fun doing your own tests on the site of Hugging Face.
Sources: Engadget, Hugging Face
4