Meta's new AI turns text and hummed tunes into music

After an AI-generated Drake and The Weeknd track blew up on the internet, the music industry went into full panic mode. Record labels immediately mobilized legal teams to scour the web for copyright infringement, proving that audio is the next major battleground for generative AI.

Now, Meta is entering the ring with a different approach. The company just introduced MusicGen, a new generative AI model that turns simple prompts into original music.

Hum a tune, get a track

MusicGen lets you generate music using text prompts, melody inputs, or a mix of both.

Basically, you can describe the exact vibe you want in text, or theoretically just hum a tune into a microphone. The system then takes those cues and outputs original variations based on its underlying instrument styles.

To make this work, Meta trained the model on 20,000 hours of music. That training set includes whole tracks alongside individual instrument samples to cover a wide range of musical styles.

While MusicGen is not widely rolled out yet, Meta has made an interactive demo available for testing.

The copyright crosshairs

The biggest question hanging over MusicGen is copyright. Music labels have already deployed task forces to hunt down unauthorized sampling and voice cloning across the web.

Yet stopping a tool like MusicGen might be much harder from a legal standpoint. Because the model synthesizes new audio from underlying structural samples rather than mimicking a specific pop star's vocals, shutting it down legally will be tricky.

Like it or not, AI-generated hits reaching the Billboard charts seems inevitable at this point.

Why it matters

For marketers, video creators, and indie game developers, MusicGen is an immediate time-saver. Sourcing royalty-free background tracks has always been tedious and expensive, and being able to generate a custom loop with a single sentence removes massive friction.

For musicians, it is not an instant replacement for a real producer, but it is an incredible rough-draft generator. If you are stuck on a bridge, humming a baseline into the tool could instantly spit back a dozen harmonic directions to explore.

We think Meta was smart to focus on instrumental styles and melodies instead of vocal cloning. Mimicking real singers is a guaranteed legal target, whereas generating instrumental arrangements lives in a much grayer, harder-to-police territory. Still, the transparency of that 20,000-hour training dataset remains an unresolved liability that could invite scrutiny down the line.