Eleven v4 for Voice Agents: What Changed From v3

Eleven v4 Turbo. Eleven v4 is ElevenLabs' new flagship voice model, built for recorded audio where quality matters more than speed: narration, voiceovers, audiobooks. Eleven v4 Turbo is the low-latency version of the same technology, built for live conversations, and it is the one ElevenLabs recommends for agents. An AI receptionist has a caller waiting on every reply, so Turbo is the right pick.

Compared with v3, both models cover more languages (90+, up from 70+), follow delivery instructions more closely, hold accents better, and clone voices more accurately. Turbo also answers faster than v3's real-time model, according to ElevenLabs' published figures.

ElevenLabs describes v4 as a new architecture rather than an update to v3. The changes that matter for a business agent:

• More languages. 90+ for both v4 models, up from 70+ for v3.
• A real-time version from day one. v4 Turbo is the low-latency model for agents; ElevenLabs lists a median inference time of about 100 milliseconds, against about 280 for v3's conversational model.
• Better delivery. ElevenLabs says v4 follows audio tags and delivery prompts more accurately, which affects how naturally it handles a reassuring or upbeat line.
• Stronger accents. Accents hold more consistently across languages, and when the agent speaks a language other than the voice's original one, it sounds native rather than carrying the original accent over.
• Better cloning. Instant clones from as little as 10 seconds of audio, closer similarity to the original voice, and support for Professional Voice Clones, the highest-fidelity option.
• Steadier over long content. Better speaker consistency across long passages, which matters more for narration than for short calls.

One thing goes the other way: ElevenLabs' help center notes that Eleven v4 does not have the Style and Speed sliders, and does not support SSML, the markup some older setups use to control pauses and pronunciation. If your v3 agent relies on either, test carefully after switching.

ElevenAgents lets you choose the voice model for each agent. These are the options a small business is likely to weigh:

From ElevenLabs' models documentation, October 2026. Latency figures are ElevenLabs' own and exclude network and application delays, so a real call takes longer end to end.

Flash v2.5 is still the fastest on paper, and for an agent that only speaks one of its supported languages it remains a reasonable choice. v4 Turbo trades a fraction of that speed for a more expressive voice and far wider language coverage. Turbo also supports audio tags and works with a pronunciation dictionary, useful for getting your business name, street names, or product names right.

The model is a setting on the agent's voice, so switching takes a minute. Testing it properly takes a little longer.

2

Open the voice settings and pick the model

In the agent's voice settings, open the model dropdown and choose Eleven v4 Turbo. The voice itself stays the same; only the model that speaks it changes.

Click to expand

The model decides how well a voice is spoken; the voice decides who your callers hear. ElevenLabs' voice library is large, and you can search and filter it before trying a voice on your agent.

Click to expand

A few things I would look for in a receptionist voice:

• Clear before characterful. A voice built for dramatic narration can tire a caller who only wants a booking. Steady, warm, and easy to follow wins on the phone.
• An accent your customers expect. A local-sounding voice feels more natural to most callers than a generic one.
• A pace that suits the job. Quick enough not to drag, unhurried enough that a caller can write down a time or address.
• Tested with your own words. Library previews use sample text. Your first message and a typical answer are the real test.

The voice is chosen in the same settings screen as the system prompt and language; step 1 of the knowledge base and guardrails guide shows where.

Yes, and v4 makes it easier and more accurate. ElevenLabs offers two kinds of clone:

• Instant Voice Clone: made almost immediately from a short recording. ElevenLabs says v4 can capture a voice from as little as 10 seconds of audio, though its cloning guide still suggests one to two minutes of clean recording for the best result.
• Professional Voice Clone: a dedicated model trained on 30 minutes or more of audio, which ElevenLabs describes as close to indistinguishable from the original. It needs a Creator plan or higher, training usually takes a few hours, and ElevenLabs verifies that the voice is yours. v4 supports these, so you get the full accuracy gains.

Click to expand

Whether you should is a separate question. A clone of the owner's voice feels personal, and regular customers may like hearing a familiar voice. But callers who recognize you may assume they are talking to you, which makes a clear "I'm the AI assistant for..." in the first message even more important. A good library voice, clearly introduced as an assistant, avoids that question entirely.

Both v4 models cover 90+ languages. ElevenLabs singles out Japanese, Brazilian Portuguese, Mandarin, and Cantonese as the biggest improvements in Turbo, and says an agent can either keep the same voice when a caller switches language or change to a different voice for a more native sound.

For a small business, the practical question is usually Spanish or French alongside English. Turn on the extra language in the agent's settings, then have a fluent speaker run a few test conversations. Listen for natural phrasing, not just correct words, and check that your knowledge base answers come through accurately in translation.

Speaking a language is not the same as being reachable in a region. A local phone number, and the rules that come with it, are set up separately from the voice; see the phone setup guide.

To hear the difference for yourself, you can try Eleven v4 Turbo on ElevenLabs with a test agent and your own first message. For running costs, see the AI voice agent pricing hub.

The bottom line: Eleven v4 is ElevenLabs' best voice model for recorded audio, and Eleven v4 Turbo brings the same technology to live calls, with more languages and a faster published response time than v3. For an AI receptionist, Turbo is the one to use, chosen in the agent's voice settings and tested with your own words before callers hear it. The agent in the corner of this page is a working ElevenLabs agent, so talk to it and judge the voice for yourself; our guide to adding an ElevenLabs voice agent to your website shows how it got there. For the rest of the lineup, head back to the ElevenLabs review.