MusicGen: Meta’s Open Source Answer to AI-Generated Tunes

·

·

In the tech landscape, one-upmanship is the name of the game and Meta has decided to play. Not content with letting Google have all the AI-powered fun, Meta has unveiled its very own music generator. And guess what? They’ve decided to open-source it.

The creation in question is MusicGen. A handy tool that takes a textual description, say, “An ’80s driving pop song with heavy drums and synth pads in the background,” and turns it into a sweet 12-second snippet of audio. If you’re feeling a bit experimental, you can even guide MusicGen with some reference audio from an existing song, and it will attempt to match the vibe and melody. Isn’t that something?

Here’s where things get interesting. Meta has fed MusicGen with an extensive diet of 20,000 hours of music. This includes 10,000 licensed music tracks and 390,000 instrumental-only tracks from ShutterStock and Pond5. They haven’t spilled the beans on the code used to train the model, but they’ve been generous enough to release pre-trained models for those tech enthusiasts with a decent GPU to play with.

So, you ask, how does MusicGen hold its own? It’s not about to put any Grammy-winning artists out of business, that’s for sure. However, the melodies it churns out aren’t half bad, especially for basic requests like “ambient chiptunes music”. Even when compared to Google’s AI music generator, MusicLM, MusicGen manages to hold its own.

But let’s get down to brass tacks. Where MusicGen really shines is in its response to complex prompts. For instance, when tasked with generating a “Lo-fi slow BPM electro chill with organic samples,” MusicGen outperformed MusicLM in terms of musical coherence, crafting a tune that could comfortably fit into any Lofi Girl playlist.

The world of generative music is evolving at a rapid pace, but it’s not without its hurdles. As MusicGen and its peers continue to ‘learn’ from existing music to produce similar effects, not all artists or AI users are comfortable with this development. The rise of homemade tracks that use generative AI to simulate authentic sounds has led to a surge in viral content. This has caught the attention of music labels who have been quick to raise intellectual property concerns.

The question of whether ‘deepfake’ music infringes on the copyright of artists, labels, and other rights holders remains murky. However, with several lawsuits currently making their way through the courts, we might not be too far from gaining some clarity on the matter.

Meanwhile, Meta has refrained from imposing any restrictions on how MusicGen can be used. They assure us that all the music used to train MusicGen was covered by legal agreements with the right holders, including a deal with Shutterstock.

Source: TechCrunch