ElevenLabs' Eleven v4 reads tone, pacing and context

ElevenLabs launched Eleven v4 and its low latency variant Eleven v4 Turbo, both built on a new architecture and both live now in ElevenAgents, ElevenCreative and the ElevenLabs API. The pitch is that these models read the line rather than just pronounce it, interpreting tone, pacing, emotion, character and context from the text.

ElevenLabs says v4 ranked number one on Artificial Analysis’ Provider Voice Arena in September 2026, and won blind head to head preference tests against Cartesia Sonic 3.6, Inworld TTS-2, and Google’s Gemini 3.8 Flash-Lite TTS and Flash TTS with roughly 75% of listeners preferring it. Those are vendor reported numbers on a leaderboard, so treat them as a strong showing rather than a settled result.

The two numbers worth separating

• Eleven v4: about 100ms median inference latency.
• Eleven v4 Turbo: about 150ms median time to first speech, against 262ms to 814ms for the models it was measured against.

These are different measurements, so Turbo being the “low latency” model while showing the larger number is not a contradiction. Turbo was optimized jointly with the ElevenAgents conversational agent platform, which is ElevenLabs’ argument that a voice model and the agent loop around it should be tuned as one system.

Control and cloning

Direction is entered in plain language or with inline tags such as [laughs], [said angrily in French accent], [light rain] and [phone buzzing], and ElevenLabs says the models follow these more accurately than earlier generations. Support for IPA phonemes was also substantially improved, which is the fix for custom pronunciation.

Voice cloning is where the practical gains are. Instant Voice Clones now need only 10 seconds of audio, Professional Voice Clones are supported, and a new method for capturing speaker identity is meant to keep a voice consistent across generations, dialogue and regenerated lines, which is what long form work in Studio and the Reader app depends on.

Multi speaker work also improves, because the model reads a whole scene and generates lines that respond to what was just said rather than stitching together isolated takes. Both models cover more than 90 languages, and a voice recorded in one now speaks others with a native accent that no longer drifts back toward the source language partway through a generation.

(Original source: ElevenLabs)