MusicGen offers the creation of music from a text. But it is distinguished by an innovative capacity, and more than promising.
You will also be interested
A soap bubble that dances to the music No need for brushes to paint with music, it’s possible with soap…
In matterartificial intelligence music generation materialartificial intelligence, we have been entitled, for the time being, to three main approaches:
- apps like Soundraw ask you to choose an atmosphereatmosphere, a musical genre, a tempo… From there, a piece is generated;
- very original, Google’s SingSong asks us to hum a piece and it then creates an accompaniment;
- Google MusicLM, Mubert or Riffusion ask us to type a “ prompt describing the music and then generate it. MusicLM seemed fairly advanced in this area, but we haven’t been able to test it yet.
It’s the 3e approach chosen by Meta/Facebook with its MusicGen. What to say ? He comes out of the operation honorably. Meta claims that the distinguishing point would be the use of a simplified generation model that was trained on 20,000 “licensed” music, and a database of 10,000 quality audio recordings.
We subjected MusicGen to various prompts (requests), including asking it to generate a “French-style” track. You can see the results in the video below. As you can see, sometimes you have to be a little detailed in your request to get the desired result.
In the manner of…
However, what distinguishes MusicGen is that it is possible to start from an existing recording. A bit like when, in MidJourney, you generate an illustration from an image that you submit to it. Here again, the video offers you two examples of pieces skilfully derived from Bach. Eventually, nothing would prevent you from charging “ Swing your what », « Roxane ” Or ” Yesterday » and to generate original pieces from these bases. However, it will be necessary to verify that no infringement of copyright is patent.
On Hugging Face
If you just want to test the beast without worrying too much about Python codePython, you can access this demo on Hugging Face. The problem is that it’s terribly slow and you have to be prepared to wait tens of minutes to generate a single song – when the app doesn’t tell you, plain and simple, that it’s saturated with requests and will have to try again later.
From Google Colab
The alternative, a priori daunting is to go through the Google Colab interface. Don’t worry, the Python code is predefined and all you have to do is click on the arrows to the left of each code block to run it. When you get to the last block of code, you can replace the prompts with queries of your choice.
