After explaining to you how to try Meta’s Llama2 AI chatbot, it’s time again to delve into what the well-known company has released in the field of artificial intelligence. This time in between is the world of music.
In this regard, as also reported by The Verge and Engadget, as well as as we learn at an official level directly from a post published on the Meta blog on August 2, 2023, what Mark Zuckerberg and his associates have decided to release open source And il tool AudioCraftwhich can generate sounds from text input.
In case you’re wondering, sure: background music can also be generated. In any case, it actually refers to a combination of three models, including one already available from a web browser through the usual official portal of the startup Hugging Face.
The model that can attract the most attention is in fact MusicGen, which allows you to generate audio using a trained AI solely through music licensed by Meta. AudioGen is instead a model trained on publicly available sound effects, which can always generate audio starting from a textual input (for example, “man speaking while other people applauding in the background“). It should be noted that there is also an improved version of the EnCodec decoder involved.
We tried to use quickly the MusicGen model (one of those related to AudioCraft) through the appropriate page of the Hugging Face official website, giving as input in English “theme song for a tech website in hip hop style“, obtaining the result that you can listen to at the bottom of the news. What is generated, usually within a few minutes (when the system is not used too much, of course), is an audio lasting just over 10 seconds. Among other things, it is also possible to provide the AI with a reference from which to “take inspiration” in terms of melody.
If you are looking for more details instead on the AudioCraft “bundle of the three models”, you might also be interested in taking a look at the further study published on the official Meta AI blog, in which there are also audio samples related to the whole. In any case, the possibility of combining the aforementioned three models in a single solution could provide a musician or an enthusiast with everything he needs to generate background music solely through AI.
Sure, Meta itself claims that the dataset the AI was trained on still lacks a certain diversity and therefore there are important limitations, but the open source approach could help to expand the horizons. For the rest, we recall that for not too long ago Google has released the AI MusicLM, although the latter is still accessible today only by researchers. But the era of AI-generated music continues to advance, step by step.