ElevenLabs adds audio reference control to Music v2

Trying to describe music with text prompts is usually an exercise in frustration. You ask for dreamy synth-pop with a melancholy edge, and the model hands you generic elevator disco. Sounds live in the ear, not a prompt box.

ElevenLabs is tackling that gap. The company just rolled out References for its Music v2 model, giving creators the ability to upload actual audio to guide their generations instead of relying purely on words.

Give it a track, not just an adjective

References lets you feed an existing audio file into Music v2, which analyzes the clip to match its musical style, instrumentation, and overall vibe.

Uploads can run anywhere from 10 seconds to 5 minutes. You can let the reference track do all the steering on its own, or pair it with a text prompt to tweak the direction, like uploading a bedroom acoustic demo and asking the AI to build an upbeat indie-rock arrangement around it.

ElevenLabs launched the feature across its entire footprint: inside ElevenMusic, in ElevenCreative for teams, and through the ElevenAPI for developers integrating audio generation into third-party tools.

The copyright safety wire

Whenever audio cloning is involved, labels and rights holders get nervous. ElevenLabs is trying to cut that problem off at the pass.

Every file uploaded through References passes through an automated copyright scan before the model generates a single beat. If the system matches the upload to an existing recording owned by another artist, the request gets blocked instantly. The company says the tool is strictly meant for musicians working with their own material or proprietary brand sounds.

Why it matters

Prompting with text was always an awkward middleman for audio. Image generation made its biggest creative leaps when image-to-image and ControlNet style references arrived, and music generation is clearly following the same trajectory.

For independent musicians and production teams, this is a massive workflow shortcut. If you already have a 30-second loop or an acoustic scratch track, uploading it gets you infinitely closer to what you hear in your head than crafting paragraphs of adjectives.

We think ElevenLabs made the right strategic move by putting copyright checks directly in front of the generator. With the music industry actively suing AI audio platforms over unauthorized training and style mimicry, launching reference-based audio without tight content verification would have painted an immediate legal target on their back.

The real test will be how aggressive that copyright scanner proves to be. If independent creators find their own demos flagged by false positives, friction will spike fast. But if the balance holds, audio-to-audio guidance looks like the new baseline for AI music engines.