Meta’s MusicGen AI tool has been trained on 20,000 hours of music

The tool can generate 15 seconds of audio, or 120 seconds for those signed up to its host platform.

Meta has created an AI music tool of its own named MusicGen, which has been trained on 20,000 hours of music.

MusicGen uses text or melody prompts to create new music, and looks to be a competitor to Google’s text-to-music AI tool, MusicLM. MusicGen can generate 15 seconds of audio based on a description provided by a user, but those signed up to its Hugging Face space will be able to access clips of up to 120 seconds.

Unlike MusicLM, MusicGen doesn’t prevent users from using artist and song names in their prompts. MusicLM on the other hand, has been trained 280,000 hours of audio clips, a much larger amount than MusicGen.

The training for the new Meta tool included 10,000 hours of “high-quality” licensed songs and 390,000 instrumental tracks, according to Meta (as reported by Decrypt).

Meta has also claimed that all the music it was trained on was “covered by legal agreements with the right holders,” however, those sources are not huge labels such as Universal Music Group, as the training set actually came from media libraries like ShutterStock and Pond5.

AI has been one of the most polarising topics in the music industry for decades, having divided musicians, music industry figures, and fans alike. A recent study found that only 17.3 percent of the producers think negatively about AI. On the other hand, 34.8 percent saw it as a force for good, while the largest group (47.9 percent) had a neutral point of view.

Despite the surprisingly low percentage of producers who saw AI in a negative light, a huge 73.1 percent of surveyed producers said they thought AI music generators could replace human music producers in the future, at least to some extent.

MusicGen is not yet widely available, though you can check out a demo here. You can also read Meta’s paper on the AI tool here.

Source: Rachel Roberts – musictech.com