Expressive synthetic voices have usually meant giving up speed. ElevenLabs says it no longer has to make that trade, with Eleven v4, a speech model out this week, and a Turbo variant that starts talking in about 150 milliseconds.
Under the hood is a redesigned architecture that ElevenLabs says picks up on a script’s context, its pacing and its tone. Delivery is steered either by tags or by writing the direction out in plain words, and names or technical terms can be pinned down phonetically.
Regenerating individual lines no longer throws a long production off, and a speaker’s delivery reflects the scene around the dialogue rather than just the one line being spoken.
One request can cover up to 10,000 characters, roughly ten minutes of audio, and longer works are broken into segments meant to hold a consistent pace; language coverage now reaches past 90, up from about 70.
A ten-second sample is enough for an Instant Voice Clone, and Professional Voice Clones are back after v3 dropped support for them.
Artificial Analysis, which runs the Provider Voice Arena leaderboard, places v4 ahead of Google’s Gemini 3.8 Flash TTS and Cartesia Sonic 3.6. Its pronunciation benchmark score is 91.7 percent, against 85.6 percent for the prior generation.
On the price sheet, v4 runs $80 per million characters and Turbo $40; through October 12 those numbers drop to $22 and $11.
