Meta unveils AI Music Generator MusicGen

Meta, the tech giant, has recently unveiled a groundbreaking AI music generator called MusicGen. According to Meta, MusicGen has undergone extensive training on a library of 20,000 hours of licensed music, making it one of the most diverse and comprehensive music models available.

MusicGen stands out due to its unique architecture. Unlike previous methods, MusicGen operates as a single-stage auto-regressive Transformer model, utilizing a 32kHz EnCodec tokenizer with 4 codebooks sampled at 50 Hz. This allows it to generate all 4 codebooks simultaneously, resulting in efficient and rapid music generation.


The main focus of MusicGen is conditional music generation, where it excels in producing high-quality music samples based on specific inputs, such as textual descriptions or melodic features. This level of control enables musicians and composers to shape the output according to their artistic vision.


| Best Mastering Plugins in 2023 for Music Producers & Audio Engineers  – Click here to checkout


In comprehensive evaluations that considered both automatic and human studies, MusicGen outperformed established music models like Riffusion, Mousai, MusicLM, and Noise2Music. These evaluations measured objective and subjective metrics, assessing the alignment between lyrics and music as well as the overall plausibility of compositions.


Furthermore, through in-depth analysis, the research team behind MusicGen has highlighted the importance of each component in the model, providing valuable insights for future improvements. The release of MusicGen by Meta is expected to have a significant impact on the music industry.


How to use MusicGen:

  1. To use MusicGen all you have to do is describe your music in the prompt
  2. You also get the option to apply a condition to the melody in the form of audio. Now what does this “Condition” mean? Using this condition You can optionally provide reference audio from which a broad melody will be extracted. The model will then try to follow both the description and melody provided.
  3. The model will then generate 12 seconds of audio based on the description you provided


Here’s a sample of what MusicGen delivered when we wrote “a chill future garage track with drums including kick, snare, percussion and a guitar melody.”



Though it wasn’t 100% close to the prompt, MusicGen did a pretty good job when it comes to making drums out of AI prompts. It is also able to distinguish between genres.


You can try out MusicGen here. MusicGen is only available to try out at the moment and is experiencing a lot of traffic since its launch. So you might have to wait for quite a while before scheduling your prompt to the server.



Image credits: Dima Solomin on Unsplash

Hemant has been actively involved in the dance music industry with over 3 years of experience. Currently working as a Project Manager and heading the We Rave You Tech division. Get in touch via email.