ElevenLabs drops Eleven v4 with 100ms Turbo voice agents

The same sentence can mean two entirely different things depending on how it is spoken. "I need you to stay calm" sounds gentle when whispered by a doctor to a patient, but urgent when shouted across a battlefield in a video game.

ElevenLabs rolled out Eleven v4, its most expressive text-to-speech model to date, alongside a low-latency sibling called Eleven v4 Turbo.

The model immediately took the top spot on Artificial Analysis's Provider Voice Arena. In blind head-to-head testing against rivals like Cartesia Sonic 3.6, Inworld TTS-2, and Google Gemini 3.8 Flash TTS, roughly 75% of listeners preferred Eleven v4.

Directing AI with bracketed stage notes

Built on a fresh architecture, Eleven v4 tackles the long-standing problem of disjointed dialogue. Instead of stitching isolated sentences together, it reads the whole scene to match rhythm, pacing, and emotional cues across multiple speakers.

The interface also lets creators act more like directors. You can instruct the model in plain language or drop inline tags directly into the script, like [laughs], [said angrily in French accent], [light rain], or [phone buzzing].

Eleven v4 sticks to these audio tags and prompts far more reliably than earlier generations. Phoneme handling via the International Phonetic Alphabet has also received an overhaul, keeping custom pronunciations on track.

Cloning is faster, too. Instant Voice Cloning now needs just 10 seconds of source audio to capture a voice, while long-form consistency and request stitching have been tightened up for audiobooks and studio narration.

Low latency meets native accents across 90 languages

Voice agents usually force a painful trade-off: pick a rich, expressive voice that lags, or pick an instant robot that sounds dead inside. Eleven v4 Turbo takes aim at that gap.

The Turbo variant clocks a median inference latency of roughly 100ms, with a median time-to-first-speech of about 150ms over WebSocket streaming. That is faster than the typical pause between two humans conversing.

Both models support over 90 languages. Crucially, ElevenLabs tackled accent drift. A voice cloned from an English speaker can now converse fluently in Japanese or Italian, adhering to native cadence rather than slipping back into an English accent midway through.

Both versions are live now across ElevenAgents, ElevenCreative, and the ElevenAPI.

Why it matters

We think ElevenLabs made a brilliant move by treating speech synthesis like scene direction rather than raw text processing. Dropping inline audio tags like sighs and background rain into the script saves indie creators and game studios dozens of hours in manual post-production.

Pushing Instant Voice Cloning down to 10 seconds lowers the barrier for global localization. Brands can now take a single founder or actor's voice and distribute high-fidelity, native-accented content across 90 markets almost instantly.

For the voice AI market, a 150ms time-to-first-speech puts immense pressure on low-latency rivals like Cartesia while raising the bar for full-stack voice assistants from Google and OpenAI, which still often default to flat, uninflected tones.

The obvious catch lies in safety. When 10 seconds of casual audio can generate high-fidelity, emotion-rich speech in dozens of foreign tongues, audio verification and anti-spoofing systems will be stretched to their absolute limits.