Piotr Dabkowski

With just a few minutes of an audio sample, ElevenLabs can clone your voice.

The unicorn start-up, co-founded by two Polish friends, Piotr Dabkowski and Mateusz Staniszewski, has developed a model that can generate a variety of realistic voices from text or pre-existing speech, and be used for dubbing, so that one can appear to sound fluent across 29 languages. 

While only two years old, the company is valued at $1.1 billion and received $80 million from Andreessen Horowitz and others in January. The startup has since partnered with several large brands including HarperCollins, the Washington Post, and TIME. 

ElevenLabs’s model is more robust, has a lower latency, and higher quality than competitors. Dabkowski expects the model to soon hit a “quality ceiling,” where the AI is indistinguishable from any human it mimics.

This poses an obvious safety challenge as it could be used to fraudulent ends, but Dabkowski says that their systems monitor usage to minimize the threat. “We can trace back every single generation, and you need to use your credit card to have access to voice cloning—so it's like using a gun that is connected to the internet,” he says. 

ElevenLabs has developed a speech classifier that can identify whether any audio was generated using their model. The company is also a member of the U.S. AI Safety Institute’s Consortium. Open-source AI audio solutions are a larger concern, Dabkowski adds. Still, some have found ways around some of ElevenLab’s safeguards, and the terrain is constantly evolving, so preventing its misuse may not be so simple.

*Disclosure: TIME and ElevenLabs have a technology agreement to create audio accessible content.