Turn Written Scripts into Acted Performances with ElevenLabs v3
Last verified: November 3, 2026
Type [whispers], [sighs] or [laughs] into your script, and ElevenLabs v3 acts out each cue at the spot where you put it. You get back a performed voice take, not a flat read. ElevenLabs built the model, officially named Eleven v3, as its emotionally rich, expressive text-to-speech engine . It first came out as an alpha in June 2025 and became generally available on February 2, 2026 .
Its audio tags come in three groups. Emotion tags include [sad], [angry] and [happily]. Delivery tags include [whispers] and [shouts]. Reaction tags include [laughs], [clears throat] and [sighs]. The model also has deeper text understanding, so it gets better stress, cadence and expressivity from the words themselves . It produces lifelike speech with a wide emotional range in more than 70 languages .
Expect to direct it, not just paste text in. ElevenLabs says the model needs more prompt engineering than its earlier models , and it rewards scripts with clear emotional context. On this page it runs as ElevenLabs V3 in text-to-speech mode and makes one finished take from each script.
What One Text-to-Speech Run Handles
What one run accepts and returns, plus how much audio a full script makes.
Script length per run
Up to 5,000 characters
Audio from a full-length script
About 5 minutes
Preset voices
26, with Rachel as default
Stability modes
Creative, Natural (default), Robust
Spoken language
Auto-detect or 71 named languages
Output per run
1 audio file
Check These Before You Generate a Take
Six checks that stop a flat delivery, ignored cues or a misread script.
Pick the Voice Before You Write Tags
The voice you choose matters more than any other setting in Eleven v3 . Pick a preset whose natural tone already suits the scene. A soft-spoken voice is a poor match for a script built around [shouts].
Set Stability for How Much You Direct
Robust makes the model less responsive to directions in the text . Use Natural or Creative for scripts with lots of tags. Save Robust for plain reads where consistency matters more than performance.
Put Each Tag Right Before Its Words
Place a tag just before the phrase it should shape. Use one or two at each turn in the scene instead of tagging every sentence. Opposing cues side by side, like [happily] next to [sad], give the performance no clear direction.
Shape Pauses with Punctuation, Not SSML
Eleven v3 doesn't support SSML break tags, and the structure of your text strongly affects the result . Use ellipses, dashes and capital letters to set rhythm and emphasis.
Choose a Language for Short or Mixed Text
Auto works out the language from your text. If your script is one short line or mixes languages, pick the spoken language yourself so Auto doesn't have to guess.
Read Enhance Prompt Suggestions Before Accepting
Enhance Prompt expands a rough idea, and every word of a suggestion you pick will be spoken. Your text only changes when you pick one, so check that it keeps your tags and any lines that must stay exactly as written.
ElevenLabs v3 or MiniMax Speech: Choose the Engine That Fits Your Script
Both engines let you write performance cues into the script itself. They differ in how many languages they cover, how much text one request takes and how they handle hard-to-read text. Each row compares what the two models do as ElevenLabs and MiniMax document them. Where a MiniMax fact applies to one release, that release is named.
| Feature/Spec | ElevenLabs v3 | MiniMax Speech |
|---|---|---|
| Release milestones | Alpha released June 2025 ; generally available February 2, 2026 | Speech-02 series released April 2, 2025; Speech-2.8 released January 23, 2026 |
| Language coverage | 70+ languages | 40 languages on speech-2.8-hd |
| Text accepted in one request | 5,000 characters, roughly 5 minutes of audio | Up to 10,000 characters per synchronous request |
| Emotion direction | Inline emotion tags such as [sad], [angry], [happily], plus delivery tags such as [whispers] and [shouts] | 7 supported emotions on speech-2.8-hd |
| Non-verbal sounds written into the script | Reactions in square brackets such as [laughs], [clears throat], [sighs] | Interjection tags in parentheses such as (sighs), on speech-2.8-hd and speech-2.8-turbo only |
| Pronunciation override | Native International Phonetic Alphabet (IPA) support | Inline IPA, Mandarin Pinyin or Cantonese Jyutping in half-width parentheses |
| Numbers, symbols and formatted text | 68% fewer errors on numbers, symbols and notation since leaving alpha | Speech 2.6 reads URLs, email addresses, phone numbers, dates and money amounts without text pre-processing |
What the Differences Mean for Your Next Voiceover
Writing performance cues into the script
Both engines take cues inside the text, so a script can carry its own stage directions. Eleven v3 uses square brackets for everything: feelings, delivery and reactions. A tagged script reads like a director's draft. MiniMax splits the job between its documented set of emotions and parenthesized sound interjections, and the interjections only work on the speech-2.8 models. If you move a script from one engine to the other, plan to rewrite the cues, because the two syntaxes don't match.
Planning around language reach and script length
Here the two engines pull in opposite directions. Eleven v3 covers far more languages, which matters when one voiceover has to ship in many markets. MiniMax's synchronous requests take twice as much text, so long documents need fewer splits. On this page, the Text box has the same 5,000-character ceiling as Eleven v3 itself. Plan a long chapter as several runs from the start.
Handling names, numbers and formatted text
Both models accept IPA for pinning down a hard name. MiniMax also documents Pinyin and Jyutping for Chinese. Their work on formatted text targets different material. Eleven v3's general release focused on numbers, symbols and notation in every language, like scores, formulas and digit-by-digit phone numbers. Speech 2.6 focused on the URLs, emails and dates that turn up in agent replies. Look at what your scripts are full of and test that kind of text first.
Choose by Language Reach or by Text per Request
How to choose: Choose ElevenLabs v3 when one engine has to cover many languages: it supports 70+, against 40 on MiniMax speech-2.8-hd. Choose MiniMax Speech when one request has to hold more text: its synchronous requests take up to 10,000 characters, against 5,000 for Eleven v3. For non-verbal sounds written into the script and IPA pronunciation, the two are equal. Eleven v3 and the MiniMax speech-2.8 models both support them.
Generate a Directed Voice Take
Four steps take you from a tagged script to a finished audio take.