ElevenLabs

Contribution to AI

Synthetic speech had been intelligible for decades without ever being mistaken for a person; the flat prosody of phone menus was the accepted ceiling. The models put into open beta in January 2023 broke it audibly — breath, hesitation and emphasis chosen from the surrounding sentence rather than the single word — and cloning a voice stopped requiring a studio and hours of tape. What followed was the field's first mass lesson in what cheap cloning does. Within days, 4chan users had made celebrities appear to say racist things, and the responses improvised that week became the industry's template: cloning restricted to identity-verified paying accounts, a free classifier that reports whether a clip came from the company's own models, and a refusal list covering political figures. When a robocall in the 2024 New Hampshire primary used a fabricated Joe Biden telling voters to stay home, the audio was traced back to those tools, which settled the argument about whether convincing impersonation was still a state-level capability. Multilingual v2, released in August 2023, carried a speaker's identity across roughly thirty languages, turning dubbing into a translation problem rather than a recasting one. The voice marketplace then attached payment to consent, letting people license their own voice and earn from each use.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).