Meta open source AI language model MusicGen can convert text and melodies into complete music

IT House News on June 12, Meta recently open sourced its AI language model MusicGen on Github, which is based on the Transformer model launched by Google in 2017. As the name of the model indicates, MusicGen is mainly used for music generation. It can convert text and existing melodies into complete music.

The R&D team stated: “We used 20,000 hours of authorized music to train the model, and used Meta’s EnCodec encoder to decompose the audio data into smaller units for parallel processing, thereby improving MusicGen’s computing efficiency and generation The speed is better than the same type of AI model.”

In addition, MusicGen also supports the combined input of text and melody. For example, you can propose to generate "a light track" and at the same time request "to combine it with Beethoven's "Ode to Joy" Combined".

The R&D team also tested the actual performance of MusicGen. The results show that compared with Google's MusicLM and other music models such as Riffusion, Mousai, and Noise2Music, MusicGen performs better in testing indicators such as the matching of music and text prompts and the credibility of the composition, and is overall slightly higher than Google's MusicLM level.

Meta has allowed commercial use of the model and has released a demo web application on Huggingface.

IT Home attaches the model address: Click here to go

Demo application address: Click here to go

The above is the detailed content of Meta open source AI language model MusicGen can convert text and melodies into complete music. For more information, please follow other related articles on the PHP Chinese website!