
Eleven v4 in Unitool Whispers, Laughs and Sighs Wherever You Put a Tag
AI voice-over often sounds as flat as a gate announcement at the airport. ElevenLabs’ Eleven v4 takes delivery cues in square brackets, and it is already available in Unitool. We show how to voice a text with a whisper and a sigh, what it costs and which ElevenLabs model to pick.
Why Your AI Voice Reads Everything Like an Airport Announcement
You wrote a warm script for a video, with a pause before the key line and a joke at the end. You pasted it into a voice-over tool and got a voice that says “20% off” and “we are sorry” with exactly the same cheer. You add exclamation points, chop sentences into pieces and rewrite the lines for the third time. The tone never changes, because there is nothing to steer it with.
On September 28, 2026, ElevenLabs released Eleven v4, a text-to-speech model whose delivery you can direct right inside the text. It is already in Unitool. First we cover what this model is and what the tests say about it. Then we show how to voice your own text with a whisper and a sigh, what it costs and which ElevenLabs model to choose for the job.
What Eleven v4 Does With Your Text
Eleven v4 turns text into speech and tries to work out how that speech should be delivered. According to ElevenLabs, it takes tone, pacing, the character of the writing and the context of the scene into account. You can hint at delivery with short notes in square brackets, called tags. The announcement gives examples such as [laughs], [light rain] and [phone buzzing], which means laughter, the sound of rain and a buzzing phone. The company says v4 follows such cues more accurately than earlier models.
Two more features are claimed. The model keeps the whole scene in mind, so in a dialogue the characters answer each other instead of reading separate lines stitched together. And a voice should stay consistent over long texts, which matters for audiobooks and ads. Support for the International Phonetic Alphabet (IPA) was also improved, so custom pronunciations of rare words and names should behave more reliably.
What the Tests Show and Why We Don’t Take Them on Faith
ElevenLabs says that as of September 2026 v4 ranks first on the Artificial Analysis Provider Voice Arena Leaderboard. In blind comparisons, according to the company, about 75% of listeners preferred it over competing models, and the win rate in individual pairs ranged from 65% to 81%. The rivals were Cartesia Sonic 3.6, Inworld TTS-2 and two Google Gemini 3.8 models, Flash-Lite TTS and Flash TTS. Listeners heard the same line from v4 and from a competitor without knowing which was which and judged which sounded more expressive and more natural. A tie counted as half a win.
A caveat is in order. ElevenLabs ran these tests itself, on its own lines and languages, so your results may differ. We do not promise that v4 will beat everything in your video. But the numbers explain why we suggest starting with it.

Eleven v4 vs. v4 Turbo
There are two models. Eleven v4 is built for voice-over where quality comes first, meaning audiobooks, narration, dubbing and characters with personality. Eleven v4 Turbo is the fast version for voice agents that answer customers in real time. ElevenLabs reports a median time to first speech of about 150 milliseconds, which is its own measurement. A video or a course does not need a delay measured in tenths of a second, so we start with regular v4. Turbo makes sense if you are building a voice bot.
90 Languages and Voice Cloning
The announcement lists more than 90 languages, compared with more than 70 for Eleven v3. A voice recorded in one language should speak others with a native accent while keeping its own timbre. According to the company, an instant voice clone can now be built from 10 seconds of audio. The Eleven v4 card in Unitool also says “90+” and lists languages such as Spanish, Japanese and Hindi. We tested it in Russian, and you will see how it went below.
Which ElevenLabs Model to Pick in Unitool
The model list in Unitool is long, and half of it is not needed for a video. We would split it into four groups.
Eleven v4 — the first choice for voice-over with emotions, 90+ languages.
Eleven v4 Turbo — the same 90+ languages, tuned for low delay, meaning live conversations.
Eleven v3 and v3 Conversational — 70+ languages, they understand tags but, as the card says, need more careful prompting than earlier models. The second one is tuned for conversational agents.
Multilingual v2, Flash v2.5 and Turbo v2.5 — the older line with 29 to 32 languages. Emotion tags do not work there, and pauses are set with the command
<break time="1.0s" />. The card describes Flash v2.5 as an ultra low latency model.
Three more v2 models, Turbo v2, Flash v2 and English v2, speak English only. For emotional voice-over we pick v4, and for a live conversation v4 Turbo. The older models fit cases where emotions are not needed.

Where to Find Eleven v4 in Unitool and What It Costs
Now the practical part. Everything happens in one window.
Open the ElevenLabs section in Unitool and choose the text-to-speech tool.
In the right panel, under Clip settings, set the Model field to Eleven v4.
Pick a speaker, paste your text and start the generation.
Take the finished voice-over with the Export button.
The whole path is a few clicks, and there is nothing to install or connect. The rate is shown right under the input field, so you can start with a single paragraph.
The field says the voice-over costs 30 tokens per 1,000 characters. On the Starter plan that works out to roughly $0.50 per 1,000 characters (our estimate from the $5 price and 300 tokens, not counting the 30 bonus tokens). If you spent the whole monthly balance only on voice, it would cover about 10,000 characters on Starter, 15,000 on Optimal and 33,334 on Advanced. The balance is shared across all neural networks, so in practice voice-over gets a part of it.

Tags in the Text and the Delivery Sliders
Tags go in English inside square brackets, right in the text of the voice-over, before the phrase they should affect. The line itself can be in any supported language. The hint inside Unitool gives these examples. Emotions are [happy], [sad], [angry], [excited], [curious] and [sarcastic]. Delivery and sounds are [whispers], [shouting], [laughs], [crying] and [sighs]. Pauses without an exact length are [short pause] and [long pause]. These are examples rather than a complete list, and the hint advises starting with a single tag.
We tried this on a Russian line with the standard Roger voice.
[excited] I’m so glad to see you! [whispers] This is our little secret.
The model played excitement at the start and dropped to a whisper at the end (we ran the Russian version of this line). One test is not a study, but it shows the main thing. Tags do not require a cloned voice, since they worked on a voice from the library.
The panel on the right has two settings that matter for v4, Stability and Clarity + Similarity Enhancement. They default to 50% and 75%. Very high stability dulls the emotion and very low stability makes delivery unpredictable, as the hint warns. The Style Exaggeration slider and the Speaker boost switch stay on screen, but they are unavailable for v4, and so is speed. v4 also ignores <break> pauses, so use tags and punctuation instead.

A Script for a 25-Second Video
Say you need to voice a short video about launching an app. Start in a Unitool text chat, for example Claude Sonnet 5.5, which writes the script rather than the voice. Here is the prompt.
Write a 25-second voice-over script for a video about the launch of an appointment-booking app for doctor visits. Keep the tone warm and free of hype. Add no more than four delivery tags in square brackets in English, for example [excited], [whispers] or [short pause], each placed before the phrase it changes. If you lack details about the app, ask me and do not make anything up.
Then copy the finished text into the text-to-speech tool, choose Eleven v4 and a speaker with a fitting timbre, and leave Stability at 50%. Listen. If a tag overdoes it, remove it and keep the punctuation, which also works. If the voice comes out flat, lower Stability and generate again. Check the app name and any numbers by ear separately, because the model may pronounce them in its own way.
How This Helps If You Are Not a Sound Engineer
Marketers can voice an ad or a short story without a studio. Make two versions of the same line, upbeat and calm, and see which one performs better.
Trainers can record a lesson video where the key point comes after a pause and the common mistake comes with light irony. Check the content yourself and leave the delivery to the model.
Small shop owners can voice a product video. One tag before the main line is often enough to stop the voice from reading like a toaster manual.
Can I Use a Famous Actor’s Voice?
The question comes up right away. No. Unitool’s rules prohibit other people’s faces and voices without consent. For a clone, upload your own recording or one from a person who agreed. The cloning page says a single file can be up to 11 MB, and you can also record a voice straight from a microphone. An account holds up to 5 voices, and unused ones are deleted 7 days after the last use.
We do not promise perfect delivery on the first try. Tags set a direction but do not guarantee an exact performance, and the hint in Unitool says so plainly. Listen to the result in full before you publish it.
Voice Your First Line
You do not have to study every model first. Take one line from your video, put two tags in it and compare it with how it sounded before.
When you sign up for Unitool, we give you 10 bonus tokens for any neural network, which covers roughly 330 characters of voice-over (our estimate). For regular work the Optimal plan at $7 a month fits, and if you use neural networks every day, take Advanced at $16 a month. Pick a plan, switch on Eleven v4 and voice your first paragraph.
At your next voice-over, what will help is not knowing every model, but the habit of putting one tag where the voice should sigh.