ElevenLabs launches Eleven v4 and a low-latency Turbo model for voice agents

A new architecture aimed at expressive speech

ElevenLabs says Eleven v4 is built on an entirely new architecture and was designed to interpret tone, pacing, emotion, character and context. The company describes output that can sound dramatic, tender, urgent, comedic or conversational while keeping the identity of the speaker. Underlying audio fidelity is also higher across the board, according to ElevenLabs, with cleaner and more natural-sounding output.

The model also targets multi-speaker scenes. Because it understands the context of a whole scene, ElevenLabs says, speakers respond to what has just been said instead of sounding like isolated lines stitched together. A new method for capturing speaker identity is meant to keep voices consistent across agent conversations, audiobooks and ads.

Users can describe how a line should be delivered in natural language. They can also use inline tags such as [laughs], [said angrily in French accent], [light rain] or [phone buzzing] to control emotion, pronunciation and sound effects. ElevenLabs says v4 follows these audio tags and direction prompts more accurately than earlier models. Support for International Phonetic Alphabet (IPA) phonemes has been significantly improved, so custom pronunciations should behave more reliably. The tag syntax and API usage are covered in the ElevenLabs developer documentation.

Eleven v4 Turbo for agents

ElevenLabs argues that high-quality voice models have tended to be slower, forcing users to choose between fast and expressive voice agents. Eleven v4 Turbo is meant to remove that trade-off. The company cites a median time to first speech of ~150ms, and also a median inference latency of ~100ms, which it says is faster than the average pause between two people talking.

The Turbo model is built to work with ElevenAgents, the company's conversational agents platform. ElevenLabs says other agent builders combine models and software from different vendors, which leaves users unable to refine outputs for their own use cases. Its research and engineering teams optimized Eleven v4 Turbo and ElevenAgents together as one system. The company points to uses such as healthcare agents that pronounce medical terms accurately and gaming agents that use slang.

The latency figure was measured in September 2026 with identical scripts and default settings, over WebSocket streaming, against Cartesia Sonic 3.6, xAI TTS, Google Gemini 3.8 Flash-Lite TTS and OpenAI GPT-4o mini TTS. Network latency was measured and removed for all systems, according to ElevenLabs.

More than 90 languages and better voice cloning

Both models support more than 90 languages. ElevenLabs says a voice recorded in one language can now speak any other fluently, taking on the accent of a native speaker while keeping the identity of the original. The company says accent adherence is noticeably stronger than before, so a voice no longer drifts back toward its source accent during a generation. For dubbing and localization, that means a chosen brand voice, such as a celebrity actor or a company tone, can be carried across supported languages.

Voice cloning has also been reworked. ElevenLabs reports significantly better speaker similarity to the source voice, and says Instant Voice Clones can now capture voices with high fidelity from just 10 seconds of audio. Eleven v4 adds support for Professional Voice Clones (PVC) for the highest-fidelity use cases. Speaker identity is preserved more reliably across generations, dialogue, narration and regenerated lines, which matters for long productions. Request stitching, used to chain generations into longer content, is also significantly more reliable, improving work in ElevenLabs Studio and the ElevenLabs Reader App.

Availability and what it means

Eleven v4 and Eleven v4 Turbo are available now in ElevenAgents, ElevenCreative and via ElevenAPI, and ElevenLabs invites users to create a free account to try them. The company positions the models for audiobooks, character performances, voiceovers, dubbing and localized conversational agents.

The ranking and preference claims come from ElevenLabs and its cited tests. The Artificial Analysis ranking is from its Provider Voice Arena Leaderboard in September 2026. The preference result is based on blind comparisons against Cartesia Sonic 3.6, Inworld TTS-2, Google Gemini 3.8 Flash-Lite TTS and Google Gemini 3.8 Flash TTS, where graders judged which clip was more expressive and which sounded more natural, with ties counted as half.

For developers building voice agents, the pitch is a single system that combines speed and emotional range. For creators, it is more controllable delivery and more consistent voices across long projects and languages.