ElevenLabs released a new voice model, Eleven v4, on 28 Sep 2026. ElevenLabs says to switch to it: it follows audio tags better than v3, and like v3 it ignores break tags. This sheet quotes ElevenLabs’ own best-practices documentation, including the parts that come with a warning. Rules that apply only to v3 are marked v3.1
- There is no prompt field. The text you submit is the performance.
- New: Eleven v4 and v4 Turbo (28 Sep 2026). v4 is the top-quality model; v4 Turbo is the fast one, for live voice agents. More than 90 languages.2
- Eleven v4 and v3 both use inline audio tags such as
[laughs]or[whispering], and ElevenLabs says neither supports SSML break tags.3 - Pauses: on v4 and v3 type
[pause]or[long pause]; on older models such as Multilingual v2 use a break tag, sparingly, because too many in one generation can make the output unstable. - Punctuation does real work. ElevenLabs says ellipses add pauses and weight, and capitals add emphasis.
- The voice still matters. ElevenLabs: “A meditative voice shouldn’t shout; a hyped voice won’t whisper convincingly.” It says v4 “can still follow a tag the voice was not trained on”, though “the result may not be optimal”. Test tags on the voice you actually use.
- Not yet documented: whether every v3 tag works the same on v4, and whether Eleven v3 will stay available. ElevenLabs had not said either on its v4 page, re-read 12 Oct 2026.
Every rule on this sheet is quoted from ElevenLabs’ own documentation or blog, linked where it is used. The copy-paste scripts are this site’s own examples.
An audio tag is a stage direction in square brackets, like [laughs], typed straight into your script. The model acts it out instead of reading it aloud. A break tag (<break time="1.0s" />) is an older, code-style way to add a pause; it belongs to SSML, a markup language for speech. ElevenLabs says v4 and v3 do not support it. A voice clone is a copy of a real voice made from recordings.
How to add a pause on each model
The plain point: on Eleven v4 and v3, type a pause as an audio tag, [pause] or [long pause], or slow the line with an ellipsis; break tags do nothing there. On older models such as Multilingual v2, use a break tag, <break time="1.5s" />, for up to 3 seconds.
| Model | How to pause |
|---|---|
| Eleven v4 | [pause] · [long pause] · ellipsis. No SSML |
| Eleven v3 | Audio tags and punctuation, as on v4. No SSML break tags |
| Multilingual v2 | <break time="x.xs" />, up to 3 s |
| Flash and Turbo models | Not stated by ElevenLabs; test it |
On v4, ElevenLabs says “SSML tags don’t work in Eleven v4” and that pauses come from audio tags such as [pause] and [long pause], whose length is not fixed: “The exact length of the pause fits the moment, rather than being a pre-defined, mechanical silence gap.” An ellipsis also works: “Adding an ellipsis will slow down the delivery of a line.” The same post shows the break tag in use on Multilingual v2.4
ElevenLabs’ docs say “Eleven v4 and Eleven v3 do not support SSML break tags”, tell you to “Use audio tags, punctuation (ellipses), and text structure to control pauses and pacing”, and say the v4 prompting techniques “also apply to Eleven v3”. Their tag-enhancing prompt lists [short pause] and [long pause] among the non-verbal tags. The docs name only v4 and v3 as lacking break tags; they do not list which other models take them, so for Flash and Turbo models this sheet found no ElevenLabs statement either way.5
Check the model first. Square-bracket tags for v4 and v3, break tags for Multilingual v2, and on any model an ellipsis is the gentlest pause.
Audio tags ElevenLabs documents
The plain point: these are the tags ElevenLabs lists by name. Others may work; these are the ones it shows. ElevenLabs warns that “Some experimental tags may be less consistent across different voices.”6
| Kind | Tags |
|---|---|
| Laughing, breathing | [laughs] [laughs harder] [starts laughing] [wheezing] [sighs] [exhales] [snorts] |
| Delivery and mood | [whispers] [sarcastic] [curious] [excited] [crying] [mischievously] |
| Pauses and small sounds | [pause] [short pause] [long pause] [clears throat] [exhales sharply] |
| Sound effects | [applause] [clapping] [explosion] [gunshot] [swallows] [gulps] |
| Experimental | [strong X accent] (X = the accent) [sings] [woo] |
| Scene sounds (v4) | [light rain] [phone buzzing] |
The Enhance button writes tags for you: ElevenLabs says “you can automatically generate relevant audio tags for your input text by clicking the ‘Enhance’ button.”7
What is new in Eleven v4
The short version: better acting from the same kind of script. You still write a script with punctuation and audio tags. v4 is meant to follow them more closely.
| What | v4 |
|---|---|
| Model IDs | eleven_v4 · eleven_v4_turbo |
| Audio tags | Yes, followed “more reliably” |
| Sound cues in tags | Yes, e.g. [light rain] |
| SSML break tags | Not supported |
| Style and Speed sliders | Not available |
| Pronunciation (IPA) | Improved support |
| Professional Voice Clones | Supported |
ElevenLabs, Eleven v4, read at source 1 Oct 2026 and re-read 12 Oct 2026: “Style and Speed sliders are not available in Eleven v4, and SSML is not supported.” · ElevenLabs, Best practices, read at source 1 Oct 2026 and re-read 12 Oct 2026: “Eleven v4 follows audio tags more reliably than earlier models, including when the voice wasn’t trained on that delivery.”; “PVCs are supported in v4”; v4 “includes improved native support for the International Phonetic Alphabet, or IPA” · ElevenLabs, Eleven v4 launch post, 28 Sep 2026, read at source 1 Oct 2026: tags such as “[light rain], or [phone buzzing]”.
IPA is the International Phonetic Alphabet, the symbols dictionaries use to show how a word sounds. v4 reads them, which helps with names and technical words. ElevenLabs also warns that audio tags on v4 are “not perfect yet”, and that an accent does not carry into another language by default.8
If you move a v3 script to v4, keep the tags and punctuation, drop any break tags, and listen again: the sliders you used on v3 are not there on v4.
The formula
ElevenLabs, Best practices, read at source 16 Sep 2026 and re-read 1 Oct and 12 Oct 2026: “Punctuation significantly affects delivery in v4”.
Template:
[Sentence with ordinary punctuation for rhythm.] [Ellipsis for a weighted pause …] [CAPITALS for emphasis.] [v4 or v3: inline audio tag] / [Multilingual v2: <break time="1.0s" />]
Worked example:
It was a VERY long day [sigh] … nobody listens anymore.
Breakdown — what each part does
The script
Ordinary prose with deliberate punctuation. This carries most of the performance. The example above is ElevenLabs' own.
Punctuation
ElevenLabs: “Standard punctuation provides natural speech rhythm”
Capitalisation
ElevenLabs: “Capitalization increases emphasis”
Audio tags (v4 and v3)
Inline square brackets for emotion, delivery and pauses. Match the tag to the voice: “Tags land more readily when the delivery is already in the voice’s training data.” Until 12 Oct 2026 this card also quoted a v3-only line, “The voice you choose and its training samples will affect tag effectiveness”, read 16 Sep 2026; it is no longer on ElevenLabs’ page, which now says the v4 advice applies to v3 too.9
Break tags (Multilingual v2; not v4, not v3)
<break time="x.xs" />, up to 3 seconds. Dashes and ellipses also pause, but ElevenLabs adds: “However, these are less consistent.”
Preparing numbers and abbreviations
The plain point: write numbers, dates, phone numbers and abbreviations the way they should be spoken before you generate. ElevenLabs says “Normalization is enabled by default for all TTS models”, but that models still read some inputs differently: Multilingual v2 “can better generalize the reading out of numbers”, while Flash v2.5 is “a much smaller model and so cannot”. It names phone numbers, currencies, calendar events, times and units (“TB” for terabyte) as the usual trouble.10
| You type | Write it as |
|---|---|
1234 | one thousand two hundred thirty-four |
555-555-5555 | five five five, five five five, five five five five |
2024-01-01 | January first, two-thousand twenty-four |
100% | one hundred percent |
Dr. | Doctor |
ElevenLabs’ fixes, in its order: use a model trained to read numbers (it notes that may not suit low-latency uses such as voice agents); if an AI writes the script, “add normalization instructions to the prompt”; “Expand all abbreviations to their full spoken forms”; or clean the text with code (regular expressions) before it reaches the voice model.7
Names and unusual words: v4 reads the International Phonetic Alphabet (IPA, see above). On older models, ElevenLabs says “Phoneme tags are only compatible with the eleven_flash_v2 model”, and suggests alias tags, or simply spelling the word the way it sounds, on models such as Multilingual v2.11
Does it take a negative prompt?
The nearest thing to a negative prompt here is restraint: fewer tags, not more.
ElevenLabs, Best practices, read at source 16 Sep 2026: “Using too many break tags in a single generation can cause instability. The AI might speed up, or introduce additional noises or audio artifacts.”
Write the script with punctuation first. Add a tag only where the plain reading came out wrong.
What not to do
[pause] instead.✕This is this site's working order, not an ElevenLabs rule: write the script with punctuation, generate, then add tags only where the plain reading failed. Tagging every line up front is the voice equivalent of a bloated negative prompt. An earlier version of this sheet said break tags on v3 "fail silently"; ElevenLabs says only that v3 does not support them, so that wording was removed. The same order works on v4.
Free or paid? What the free plan allows, who owns the output, and when you need to pay: free vs paid AI image, video and music tools.
Where this model sits among the others: which models take a negative prompt, and what each vendor publishes as its formula.
Voice and music prompting more generally is on prompting music and voice.
The checklist
Before you send it:
[pause] on v4 and v3, break tags of 3 seconds or less elsewhere?□ElevenLabs, Text to Speech best practices, read at source on 16 Sep 2026 and re-read 1 Oct 2026. Added 1 Oct 2026: ElevenLabs, Eleven v4 (docs) and Eleven v4 launch post, 28 Sep 2026, both read at source 1 Oct 2026. Added 12 Oct 2026: ElevenLabs, How to add pauses in AI voiceover with Eleven v4, 2 Oct 2026, read at source 12 Oct 2026; best-practices and v4 pages re-read the same day.
Model versions and vendor documentation both move. Re-read the source before relying on a specific number.
Every other sheet in this set: all model cheat sheets.
Related on this site: ElevenLabs Alternatives · What Are ElevenLabs Skills? · How to Make ElevenLabs Voices Sound Natural · How to Write an ElevenLabs Voice Design Prompt
- 1ElevenLabs, Best practices, re-read at source 1 Oct 2026“In general, Eleven v4 is a net upgrade over Eleven v3, delivering better results in almost every case.” ↑
- 2ElevenLabs, Eleven v4 launch post, 28 Sep 2026, read at source 1 Oct 2026v4 Turbo has “a median time to first speech of ~150ms” (ElevenLabs’ figure); “More than 90 languages”. ↑
- 3ElevenLabs, Best practices, read at source 1 Oct 2026“Eleven v4 and Eleven v3 do not support SSML break tags.” ↑
- 4ElevenLabs, How to add pauses in AI voiceover with Eleven v4, 2 Oct 2026 (updated 3 Oct 2026), read at source 12 Oct 2026.↑
- 5ElevenLabs, Best practices, read at source 12 Oct 2026“Use <break time="x.xs" /> for natural pauses up to 3 seconds.” ↑
- 6ElevenLabs, Best practices, read at source 12 Oct 2026 (groups and spellings as listed) · ElevenLabs, How to add pauses in AI voiceover with Eleven v4, 2 Oct 2026, read 12 Oct 2026 (
[pause],[long pause]) · ElevenLabs, Eleven v4 launch post, 28 Sep 2026, re-read 12 Oct 2026 (sound cues).↑ - 7ElevenLabs, Best practices, read at source 12 Oct 2026.↑a↑b
- 8ElevenLabs, Eleven v4, read at source 1 Oct 2026.↑
- 9ElevenLabs, Best practices, read at source 1 Oct 2026 and re-read 12 Oct 2026 (v4 line; the v3 line read 16 Sep 2026 is no longer on the page).↑
- 10ElevenLabs, Best practices, “Text normalization” section, read at source 12 Oct 2026.↑
- 11ElevenLabs, Best practices, “Pronunciation” section, read at source 12 Oct 2026.↑
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK