ElevenLabs audio tags: how to add emotion, pauses and whispers
Short answer: ElevenLabs audio tags are short bracketed instructions you type directly into your script, like [whispers] or [sighs], that tell the Eleven v3 and v4 Turbo models how to perform a line rather than just what words to say. As of October 2026, there's no support for formal SSML break tags in v3 or v4, so pauses are controlled with audio tags, punctuation and line structure instead of a timed break tag. Tags work best used sparingly, one or two per line, placed right before the word or phrase they should affect.
You want your ElevenLabs script to actually sound like a performance, not a flat read, and you want the real tag syntax instead of guessing.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
This page may contain affiliate links from which Blur Studio earns commissions.
ElevenLabs is a text to speech service: you type or paste a script, pick a voice, and it generates spoken audio from it. Its newer models, Eleven v3 and Eleven v4 Turbo, add a feature called audio tags, short bracketed words typed directly into your script that change how a line is performed, not just what it says. This guide covers the real tag syntax, how to control pacing and pauses without relying on SSML, and the mistakes that most often ruin a generation.
What are ElevenLabs audio tags?
An audio tag is a word or short phrase in square brackets, placed inline in your script, that the model reads as a performance instruction instead of as text to speak out loud. Write [whispers] "Don't tell anyone I said this." and the model lowers its voice to a whisper for that line, rather than literally saying the word "whispers." This only works reliably on the newer v3 and v4 Turbo models. Older, faster models built for real-time use generally don't interpret tags the same way, since they're optimized for speed over expressive range.
The tag list, with real examples
ElevenLabs groups tags into a few rough categories. These are documented examples, not an exhaustive list, the model also responds to tags you invent yourself if they're a plausible emotion or sound, but the ones below are the reliable, documented starting point.
- Emotion and delivery: [excited], [whispers], [sarcastic], [curious], [crying], [mischievously]. Example: [curious] "Wait, how did you know that?"
- Laughter and non-verbal reactions: [laughs], [laughs harder], [starts laughing], [wheezing], [sighs], [exhales], [snorts]. Example: "I told him the deadline was Friday." [sighs] "He heard me."
- Sound effects dropped into dialogue: [gunshot], [applause], [clapping], [explosion]. These are meant for short dramatic reads, not business narration, and are easy to overuse by accident.
A practical customer service example from ElevenLabs' own documentation shows tags set the tone for an entire opening line: [professional] "Thank you for calling Tech Solutions. My name is Sarah, how can I help you today?" One tag at the start of a line can carry the tone for the whole sentence, you don't need to tag every word.
Tags and voice settings work together
Audio tags aren't the only lever over delivery. ElevenLabs' voice settings panel, available alongside the voice picker, carries sliders for stability, similarity and style that interact with how strongly a tag actually comes through. A stability slider pushed high tends to flatten a voice toward a consistent, even delivery, which can mute an emotion tag's effect, since the model is being told to stay close to its default read regardless of instruction. Pushed lower, the voice has more room to vary, which generally lets tags like [excited] or [whispers] land more noticeably, at the cost of some consistency across a long script. If a tag you're using doesn't seem to be doing much, check the stability setting before assuming the tag itself isn't working, it's a common reason a tag reads as weaker than expected.
For most business narration, I'd rather start with a moderate stability setting and a light touch on tags, then adjust one variable at a time, stability first, then tag choice, rather than changing both at once and losing track of which change actually fixed the read. ElevenLabs voice settings explains stability, similarity and style one by one.
How to control pauses without SSML
If you've used other text to speech tools, you might expect a timed pause tag like <break time="500ms" /> from SSML, the standard markup language several speech APIs support. ElevenLabs is explicit that Eleven v3 and Eleven v4 do not support SSML break tags. Instead, pacing and pauses are controlled three other ways: Adding pauses and whispers to an AI voiceover covers more cases.
- Punctuation: a comma, period or ellipsis each create a different length pause. An ellipsis ("Well... I'm not sure.") reads as a longer, more hesitant beat than a comma.
- Line breaks and paragraph structure: splitting a script into shorter lines, the way you'd structure a script for a human voice actor, tends to produce more natural pacing than one dense paragraph.
- Audio tags that imply a pause: [sighs] or [exhales] placed between two clauses naturally creates a beat without a formal break tag.
In practice, this means pacing a script for ElevenLabs is closer to writing a screenplay than writing markup. Read your script out loud yourself first. Wherever you naturally pause, that's where a comma, a line break or a short tag belongs. Why AI voice changes tone in long audio explains drift over long scripts.
Common mistakes with audio tags
- Stacking too many tags in one line. [excited] [whispers] [sarcastic] in a single sentence confuses the model more often than it helps. Pick the one emotion that actually matters for that line.
- Using tags on an older, non-v3 model. If your account defaults to a fast, low-latency model for real-time use, the bracketed text can get read out literally as spoken words instead of being interpreted as an instruction. Confirm you're generating with Eleven v3 or v4 Turbo before troubleshooting anything else.
- Expecting an SSML break tag to work. It won't on these models. If a pause doesn't land the way you want, change the punctuation or line structure first, not the tag.
- Forgetting to remove a tag that leaks into the audio. ElevenLabs' own documentation notes that descriptive text can sometimes get spoken out by the model and need manual trimming afterward. Always listen to the full generation before publishing, not just the first few seconds.
Where this fits into a real script
For a training video, an ad or a phone script, I usually tag sparingly: one emotional cue at the start of a section, maybe a [sighs] or a short pause where a human presenter would naturally breathe, and nothing more dramatic than that. Business narration doesn't need [gunshot] or [laughs harder], those are built more for character work and fiction. If you're testing this yourself, generate the same line three ways, no tags, one tag, and a punctuation-only version, and listen back before deciding which actually sounds most natural for your use case. You can test this directly inside ElevenLabs' text to speech tool. Making an AI voice sound less flat in ad scripts has examples.
How I'd use this for a client
When I build a voice feature into a client's product, an IVR phone tree, an automated training narrator, a product demo voiceover, getting the pacing right usually matters more than picking the fanciest voice. I script it the way a voice actor would read it, test tags against a punctuation-only version, and only keep what's actually audible as an improvement rather than tags for their own sake. Get in touch if you want a voice feature built properly around your actual script, not guessed at after the fact.
Questions people ask
Do audio tags work on every ElevenLabs model?
No. They're built for Eleven v3 and Eleven v4 Turbo. Faster, low-latency models used for real-time conversation generally don't interpret tags the same way, and the bracketed text can get read out literally instead.
Can I use SSML break tags for pauses?
Not on v3 or v4. ElevenLabs states these models do not support SSML break tags. Use punctuation (commas, periods, ellipses), line breaks, and audio tags that imply a pause instead.
Can I invent my own audio tags?
The model will often attempt to interpret a plausible tag you write yourself, but the documented list (emotion tags like [whispers] or [excited], and non-verbal tags like [sighs] or [laughs]) is the reliable starting point to test from.
Why did my tag get spoken out loud instead of performed?
This usually means you're generating on a model that doesn't support audio tags, or the tag wasn't recognized. Confirm you're using Eleven v3 or v4 Turbo, and listen to the full generation, since ElevenLabs' own documentation notes descriptive text can occasionally leak into the audio and need manual trimming.
How many tags should I use per line?
Usually one, placed at the start of the phrase it should affect. Stacking several emotion tags in one sentence tends to confuse the read rather than improve it.
Are audio tags useful for business narration, not just character voices?
Yes, used sparingly. A single tone-setting tag at the start of a line, plus natural punctuation for pacing, works well for training videos, ads and phone scripts. The more dramatic sound-effect tags are built more for fiction and character work.