The difference between a voiceover that plays and a voiceover that performs is in the script — specifically, what you put in brackets.
You've generated a voiceover in ElevenLabs. The voice is technically good — clear pronunciation, natural pace, reasonable tone. But it sounds like someone reading from a teleprompter. Flat. Professional. Forgettable.
What you wanted was a character. Someone who sounds scared before they admit what happened. Someone whose voice tightens with anger before they say the line you wrote. Someone who whispers the most important sentence of the script because that's how it should be delivered.
The tool that exists to give you exactly that is ElevenLabs' audio tag system — and most people using ElevenLabs have never touched it.
On June 5, 2025, ElevenLabs released Eleven v3 Alpha and introduced audio tags: short text commands placed inside square brackets that the model interprets as performance cues rather than text to be spoken. Unlike previous text-to-speech models that simply read text aloud, v3 performs it with situational awareness and emotional intelligence.
This guide is the complete breakdown of how to use audio tags — combined with punctuation techniques that work across all ElevenLabs models — to transform a flat read into a dramatic performance.
What Audio Tags Actually Are
ElevenLabs Audio Tags are words wrapped in square brackets that the new Eleven v3 model can interpret and use to direct the audible action. They can be anything from [excited], [whispers], and [sighs] through to [gunshot], [clapping] and [explosion]. Audio Tags let you shape how AI voices sound, including nonverbal cues like tone, pauses, and pacing.
The critical distinction: tags in square brackets are never spoken alowa. They are stage directions. The voice reads the text around them and performs according to what the tags instruct. This means you can write a script that looks like a screenplay — performance cues embedded directly alongside dialogue — and ElevenLabs will execute it.
Examples of voice delivery tags include: [whispers] I have to tell you a secret. [angry] That was never the plan.
In the first example, the word "whispers" is never spoken. The voice simply whispers the sentence that follows. In the second, "angry" is never spoken — but the sentence that follows it is delivered with audible frustration and edge.
That's the core mechanic. Everything else is craft.
The Four Tag Categories You Need to Know
Tags fall into four functional categories. Understanding what each one does helps you combine them intelligently rather than guessing.
1. Emotional Tags — Setting the Core Feeling
The core emotional tags include [happy] for expressing joy and positivity, [excited] for adding enthusiasm and energy, [sad] for conveying melancholy, [angry] for expressing frustration, [nervous] for showing anxiety, [curious] for indicating inquiry and interest.
The complete emotional tag library includes:
Positive emotions: `[happy]`, `[excited]`, `[cheerfully]`, `[playfully]`, `[joyful]`
Negative emotions: `[sad]`, `[angry]`, `[sorrowful]`, `[frustrated]`, `[annoyed]`, `[flustered]`
Neutral/complex states: `[nervous]`, `[tired]`, `[calm]`, `[curious]`, `[resigned]`, `[deadpan]`, `[flatly]`
Emotional tags placed at the beginning of a line set the tone for the entire line. Placed mid-sentence, they allow dynamic shifts within a single delivery — the voice can start one way and pivot.
2. Delivery Tags — Controlling Volume and Intensity
These tags control how something is said rather than what's felt when it's said.
These are more about the tone and performance. You can use these tags to adjust volume and energy for scenes that need restraint or force. Examples include: [whispers], [shouts] and even [x accent].
Delivery tags in practice:
- `[whispers]` — intimate, close, low-volume. Essential for secrets, confessions, horror
- `[shouts]` — full-volume delivery. Use sparingly; it works precisely because of contrast
- `[softly]` or `[gently]` — quiet but not whispered. Warm, careful, tender
- `[slowly]` — deliberate pacing, weight on each word
- `[urgently]` — clipped, fast, pressured
3. Reaction Tags — Human Moments
True natural speech includes reactions. You can use this to add realism by embedding natural, unscripted moments into speech. For example: [laughs], [clears throat] and [sighs].
The model can produce [laughs] or [laughing] for natural laughter, [sighs] for resignation or tiredness, [crying] for deep sadness, [clears throat] for attention-getting sounds, [gulps] for nervous reactions, and [gasp] for surprise or shock.
Reaction tags are the ones that make AI voiceover genuinely surprising. A sigh before a difficult confession. A gulp before something frightening. A laugh that breaks through mid-sentence. These are the sounds of a person, not a performance, and they add a texture to voiceover that no amount of emotional tagging achieves on its own.
4. Cognitive Beat Tags — Pacing and Hesitation
Cognitive beats: [pauses], [hesitates], [stammers], [resigned tone].
These tags control the rhythm of thought in a performance. A hesitation before delivering bad news. A stammer when a character is caught off guard. A pause that lets a line land before moving forward. They're the difference between dialogue that sounds rehearsed and dialogue that sounds lived-in.
The Punctuation Layer — Works on All ElevenLabs Models
Audio tags are v3-specific. But punctuation-based styling works across all ElevenLabs models, including older versions you may be using for consistency with established voice clones.
Periods ".": add restraint and finality. Ellipses "...": convey thoughtfulness or hesitation. Exclamation marks "!": add energy or excitement.
Here's how each punctuation mark functions as a performance tool:
Ellipsis (...) creates a natural pause — a moment of hanging in the air, of thought unfinished, of breath held. Use it where you want the voice to slow down and let silence carry weight.
Without: "I don't know what to say."
With: "I don't know... what to say."
The second version has a pause mid-thought that communicates the character searching for words — uncertainty performed, not described.
CAPITALIZATION signals emphasis. A capitalized word receives noticeably more stress and volume than surrounding words. Use it for the word that carries the emotional center of a sentence.
Without: "That was never the plan."
With: "That was NEVER the plan."
The second version puts audible stress on "never" — the word that carries the character's anger and conviction.
Exclamation marks (!) increase energy. Multiple exclamation marks (!) increase it further, though with diminishing returns. Use sparingly or the voice sounds manic rather than energetic.
Commas create micro-pauses. Inserting a comma where one wouldn't grammatically belong adds a brief breath that changes the rhythm of delivery. Used carefully, it creates a sense of weighted speech — each phrase considered before it's spoken.
Without: "You need to leave right now."
With: "You need to leave, right now."
The comma after "leave" creates a fractional pause that makes the line feel like a warning rather than an instruction.
Before and After: The Same Script, Two Versions
The fastest way to understand what this system does is to see the same text treated two different ways.
The flat version — no tags, no punctuation styling:
"I didn't think you would actually show up. I've been waiting for a long time. We need to talk about what happened that night. Some things were said. Some things were done. And I think you owe me an explanation."
Read that aloud in your head. It's clear. It covers the content. It sounds like someone reciting text.
The styled version — audio tags and punctuation combined:
"[surprised] I didn't think you would actually show up. [pause] I've been waiting... a long time. [tired] We need to talk about what happened that night. [slowly] Some things were said. Some things were done. [angry] And I think you OWE me an explanation."
Now read that aloud. The surprise at the opening sets an emotional starting point. The pause and ellipsis create a beat of quiet control. "Tired" shifts the delivery to something heavier, more depleted. "Slowly" signals that the next lines should land one at a time, each with weight. The final line's emotional tag plus capitalized "OWE" delivers a controlled burst of anger on the sentence that needs it most.
Same words. Completely different performance.
Scripting Specific Emotional States
Here's how to approach some of the most common emotional performances you'll need to produce.
Deep Grief or Sadness
Tags: `[sad]`, `[crying]`, `[sighs]`, `[slowly]`
Punctuation: ellipsis throughout, commas to slow pacing, no exclamation marks
Example:
> [sad] She used to call every Sunday. [sighs] I keep picking up the phone... and then I remember. [slowly] There's no one there anymore.
Controlled Anger
Tags: `[angry]`, `[frustrated]`
Punctuation: CAPITALIZED emphasis words, short sentences, periods for finality
Example:
> [angry] I gave you every chance. I TOLD you what would happen. [pause] This is on you. Not me. You.
Controlled anger is more chilling than shouting. Short sentences. Full stops. The voice that doesn't need to raise itself to make the point.
Building Rage — The Escalation
Tags: start with `[frustrated]`, escalate to `[angry]`, end with `[shouts]` for maximum impact
Punctuation: ellipsis to tension-build, then capitalized words at the peak
Example:
> [frustrated] I asked you once. [pause] I asked you TWICE. [angry] How many times do I have to say the same thing before you actually LISTEN? [shouts] ENOUGH.
The tag sequence creates an arc — frustration, then anger, then explosion — that mirrors real human escalation. The `[shouts]` tag on a single word at the end lands far harder than shouting the whole passage.
The Whispered Confession
Tags: `[whispers]`, `[nervous]`, `[hesitates]`
Punctuation: ellipsis, lowercase only (avoid caps — whispers are quiet, not emphatic)
Example:
> [whispers] I need to tell you something. [hesitates] And you can't... you can't tell anyone. [nervous] Promise me. Promise me first.
Whispered dialogue works best when you also control the surrounding text's rhythm. Short sentences. Hesitations that feel real. The "promise me" repetition in the example above creates an emotional beat that rewards the `[nervous]` tag set before it.
Laughter Breaking Through
Tags: `[laughs]`, `[laughing]`, `[between laughter]`
Punctuation: exclamation marks, interrupted sentences
Example:
> [laughing] I can't — I can't even explain it. [between laughter] He just walked in! Completely straight-faced! [laughs harder] As if nothing happened!
This example shows how fast audio tags can transform a flat read into something dynamic, textured, and genuinely funny. When you push performance direction into the prompt, Eleven v3 responds with a level of expressiveness that feels recorded, not generated.
Fear and Dread
Tags: `[nervous]`, `[gulps]`, `[whispers]`, `[hesitates]`
Punctuation: ellipsis for hanging sentences, lowercase to suppress energy
Example:
> [nervous] I heard it again last night. [pause] That... sound. [gulps] Whatever it is, it's getting closer. [whispers] I think it knows I'm here.
The `[gulps]` reaction tag in the middle of a fear sequence is one of those moments that makes a listener physically react. It's not just a performance — it's a body responding to fear, and the model renders it with remarkable authenticity.
Layering Tags for Complex Performances
Place tags logically within your script. Tags at the beginning of a line set the tone for the entire line; placing them mid-sentence allows dynamic shifts. Layer tags for complexity. Combine emotional, delivery and accent tags to guide multifaceted performances. The v3 model understands layered instructions and modulates accordingly. Use reaction tags sparingly to avoid clutter. While natural speech includes sighs and laughs, overusing them can distract listeners. Use them to highlight key moments or character traits.
Layering means placing multiple tags in sequence for compound effects. Some combinations that work well:
`[tired][slowly]` — exhausted, deliberate speech. Someone at the end of their rope saying something that costs them.
`[nervous][whispers]` — frightened and quiet. Classic horror or confession register.
`[angry][shouts]` — pure explosion. Use once per script for maximum impact.
`[sad][hesitates]` — grief with difficulty speaking. The voice catches on words.
`[excited][laughs]` — joy breaking through. Genuine delight performed, not described.
Tags should flow naturally through the script — appearing at the exact moment the emotion arrives in the narrative, not front-loaded at the top of every sentence. The script should read like a director's annotated performance note, not a list of instructions.
Practical Tips for Getting Consistent Results
Use v3 for tagged scripts. Audio tags are a v3-exclusive feature. If you generate a tagged script on an older model, the tags will be read aloud as text, not interpreted as performance cues. Confirm you're on the Eleven v3 model before generating any tagged voiceover.
Match voices to tags. Professional Voice Clones (PVCs) may not yet be fully optimized for v3; ElevenLabs recommends using Instant Voice Clones or designed voices to ensure accent and emotion tags render properly.
Generate multiple variations. Emotional performances from v3 are somewhat variable — the same script can produce noticeably different emotional intensities across generations. Generate two or three versions of high-stakes lines and pick the best take.
Don't over-tag. Proper punctuation significantly impacts audio tag effectiveness. Ellipses create natural pauses, capital letters add emphasis, and standard punctuation helps establish rhythm. A script with a tag on every sentence loses the contrast that makes individual tags effective. Silence and neutral delivery make emotional moments hit harder precisely because they're not surrounded by other emotional moments.
Test tags on short passages first. Before scripting a full three-minute voiceover with complex tag layering, test your intended tag combinations on two or three sentences. Confirm the model is interpreting them the way you intend before building an entire performance around an assumption.
Use ellipsis and capitalization even on older models. These punctuation techniques predate v3 and work across all ElevenLabs models. If you're using a v2 voice clone for consistency, punctuation styling is your primary tool for expressive delivery — and it's more powerful than most people realize.
What This Unlocks for Creators
The practical applications of expressive ElevenLabs voiceover extend across almost every content format.
Audiobooks and fiction narration — character voices with distinct emotional registers, dialogue that sounds performed rather than read, tension that builds authentically across a scene.
Faceless YouTube content and documentaries — narration that grips rather than informs. The difference between a viewer watching to the end and clicking away at the two-minute mark is often whether the voice creates any emotional pull.
Podcast production — scripted segments with hosts who sound present and engaged, not robotic. The `[laughs]` tag alone changes the texture of a scripted podcast segment completely.
Game and interactive media dialogue — character lines with emotional depth that responds to the in-game moment, without recording booth sessions for every variation.
Short film and video voiceover — dramatic narration that rivals professional voice acting, produced in minutes rather than hours of recording, directing, and editing.
The monotone AI voice is a solved problem. The tool exists. The tags are documented. The punctuation techniques are learnable in an afternoon.
What you do with a voice that can actually perform is the interesting part.
What's the most emotionally complex voiceover you've tried to produce in ElevenLabs? Drop the tag combination that worked — or the one that surprised you — in the comments below.
0 Comments