ElevenLabs launches v4 speech model with 150ms latency, tops Voice Arena leaderboard

ElevenLabs has released Eleven v4, a new text-to-speech model the company says follows performance cues such as laughter and whispering more accurately than earlier versions and keeps a speaker's voice consistent across lengthy productions. A Turbo variant of the model begins generating speech in 150 milliseconds, a latency figure aimed at real-time voice agents rather than pre-recorded narration. The launch puts the New York-based startup's flagship model directly against rivals Cartesia and Google's Gemini on the third-party Artificial Analysis Voice Arena leaderboard, where v4 currently sits in front.

The expressiveness claim is narrower than it sounds, and that is the point. Laughter and whispering are the cues that most often break synthetic speech: models tend to flatten them into neutral delivery or produce artifacts that pull the listener out of the performance. ElevenLabs says v4 tracks those instructions more faithfully, which matters most for audiobook narration, dubbing and character work where a single misread line can force a re-record of a whole session.

Voice consistency over long runs is the second half of the pitch, and arguably the harder engineering problem. A narrator who drifts in pitch or timbre across a nine-hour audiobook is unusable no matter how well any single sentence sounds, so producers have historically stitched together short generations and hoped the seams do not show. A model that holds identity across an entire production removes a step that has kept AI narration out of higher-end publishing workflows.

The 150-millisecond Turbo figure is the number to watch for the agent market. Conversational voice agents live or die on turn-taking latency, and anything approaching a full second of delay makes an exchange feel like a walkie-talkie call. At 150 milliseconds, ElevenLabs is positioning v4 Turbo for phone support, in-app assistants and other live interactions, a segment where the company already sells heavily and where OpenAI, Google and a cluster of smaller vendors are competing on speed as much as voice quality.

The leaderboard placement gives the launch an external reference point rather than a vendor benchmark. Voice Arena aggregates blind listener comparisons, so a ranking above Cartesia and Gemini reflects human preference on the clips tested, not a lab measurement. That is a meaningful signal for buyers comparing vendors, though leaderboard position can shift as competitors ship updates and as the tested prompt set changes.

What to watch next is whether the expressiveness gains survive outside curated demos. Laughter, whispering and long-form consistency are exactly the cases where demo clips look best and production workloads expose failure modes, so the real test will come from audiobook publishers, dubbing houses and agent developers running v4 at volume. Pricing, availability across ElevenLabs' tiers and how quickly Cartesia and Google respond on latency will shape whether the ranking holds through the next quarter.