Skip to content
Log inSign up free

Prompting guide

How to prompt Eleven Speech

Eleven Speech prompt guide: how to direct ElevenLabs v4 with audio tags like [whispers] and [laughs], pauses, pronunciation, and voiceover examples.

Try Eleven v4Eleven Speech model page

Quick start

Your first great Eleven Speech prompt

Everything a beginner needs in one minute: what to know first, the rules that matter most, a template to fill in, and prompts to copy.

What to know first

Your prompt is the script. Eleven Speech speaks every word you write, in the voice you pick, except the words in square brackets: those are audio tags, and they direct the performance, as in [whispers], [laughs], [sarcastic], or [sighs]. Write the lines the way they should sound, put a tag right before the words it colors, and choose a voice that suits the part. The Voice setting is the character, and the clip lasts as long as the words take to say.

The rules that matter most

  1. Write only what should be said

    Everything outside square brackets is spoken, stage directions included. "She said softly" is read aloud; [softly] is performed.

  2. Put a tag before the words it colors

    [annoyed] This is hard. A reaction goes right after the words: This is hard. [sighs]

  3. Tag only what you can hear

    Emotions, ways of speaking, and sounds a voice makes: [excited], [whispers], [chuckles]. A voice cannot perform [smiling] or [walks away].

  4. Pick a voice that fits the tags

    A delivery the voice already has lands best. A soft, warm guide will not shout convincingly, and a polished anchor may ignore [giggles]. Choose the voice first, then tag.

  5. Write numbers and symbols as words

    "Forty-five dollars", not "$45"; "two thirty PM", not "14:30". The voice can misread digits, symbols, and abbreviations.

Fill in this template

Replace each part in square brackets with your own idea, and leave out a part you do not need.

Template
[Direction tag] [First line]
[Delivery tag] [Next line] [Reaction tag]
The template, filled in
[Warm, conversational tone] I wasn't going to tell anyone this... but I've been practicing.
[whispers] Every night. In the garage. [laughs]
  • [Direction tag] How the whole passage sounds, kept in its square brackets: [Warm, conversational tone].
  • [First line] The words to speak, written for the ear with full punctuation, and no brackets.
  • [Delivery tag] How the next words sound, in square brackets: [whispers], [excited], [sarcastic].
  • [Next line] The next words to speak.
  • [Reaction tag] A sound the voice makes after the words, in square brackets: [laughs], [sighs], [clears throat]. Leave it out if the line needs none.

Prompts to start from

  • One line, three moods

    [appalled] Are you serious? [sighs] I can't believe you did that!

    Why it works: A tag sets the mood before the words, a second one reacts between them, and the punctuation carries the rest.

  • A product voiceover

    [excited] Meet the lightest trail shoe we've EVER made. [short pause] Two hundred grams... [impressed] and it still grips wet rock. [warm, confident] Go further. Feel less.

    Why it works: Capitals stress one word, the weight is written out in words, and a pause tag and an ellipsis set the rhythm. Set the Voice to Elowen.

  • Character acting

    [Low, steady voice, restrained urgency] Keep the lantern covered. If they see the light, they'll know we crossed the river.

[Quietly, with controlled fear] I heard them at the bridge... not soldiers. Something else.

[Voice rising into firm resolve] Then we don't stop. We reach the tower before sunrise.

    Why it works: Each beat opens with a note to the actor in plain words, and a blank line separates the beats. Set the Voice to Sawyer and move Stability toward Expressive.

  • Another language

    [cheerfully] ¡Buenos días! Hoy vamos a preparar el mejor café de olla de tu vida... [whispers] con un ingrediente secreto.

    Why it works: Write the text in the language you want spoken; every voice speaks it fluently, and the tags can stay in English.

Advanced

Advanced Eleven Speech prompting

Techniques for more control once the basics work.

Choosing a voice and Stability

Eleven Speech has no separate place to describe a voice and does not copy one from a clip: the Voice setting is the character. Read the Voice menu like a casting list and pick the voice whose natural range fits the part. If a tag does not land, rephrase it as a quality of the voice ([low, gravelly voice]) or switch to a voice that already has that delivery.

Stability sets how much the read moves. Toward Expressive, the voice follows tags more boldly and varies from take to take; toward Steady, it stays close to its own baseline. Try about 0.3 for acted, tag-heavy lines and about 0.7 for narration that must sound the same across takes.

  • Narration and storytelling: Darian, Elara, Florence, or Sawyer. Upbeat ads and explainers: Elowen, Jade, or Lawrence. Casual, friendly lines: Maisie, Warren, or Baxter. News-style reads: Alicia or Eldrin. A guide or a mentor: Talia, Caleb, or Wyatt. A game hero: Kaelen.
  • The same Seed gives a similar take, not an identical one. For an important line, generate two or three takes and keep the best.

Structure: write for the ear

How the text is written shapes the read more than anything else. Write natural spoken sentences with full punctuation, and put a line break between beats. Then add tags, without changing a word of the script.

For long narration, open each paragraph with one direction tag for the whole passage, written like a note to an actor: [Quiet, reflective narration], [Building tension, measured pace], [Warm, intimate narration]. Then tag only the moments inside it that change.

  • Emotion goes in a tag, never in narration around the line: "she asked, her voice trembling" is read aloud, while [trembling] Are you leaving? is performed.
  • When the story itself says "He laughed loudly", keep those words and add [chuckles] after them if you want the sound too. Never turn the narration into a tag.
  • Vary the delivery across lines (energetic, relaxed, surprised, thoughtful), and never choose a tag that contradicts what the line means.
  • Speech at a normal pace runs about 2.5 words a second, so a 30-second voiceover holds about 75 words. To fit a shot, trim the script rather than rushing the read.

Long narration, a tag per passage

The rain had stopped before dawn, leaving the old road silver beneath the moon.

[Quiet, reflective narration] She had promised herself she would not return. Yet there she was, at the gate, with the key cold against her palm.

[Building tension, measured pace] The lock turned easily. Too easily. [Softly, with wonder] On the table sat a single lantern... already lit.

Why it works: The first paragraph reads in the voice's own style, each later one opens with a note for the whole passage, and one tag inside marks the turn. Set the Voice to Florence.

Audio tags for emotion and delivery

An audio tag is a word or a short phrase in square brackets that the voice performs instead of saying. Put it directly before the words it colors, directly after them for a reaction, or at a natural pause. Tags combine ([whispers] [nervously]), and you can write your own in plain words: [under his breath], [trying not to laugh], [voice breaking].

Eleven Speech makes sound effects as well as voices, so a short, vague tag can be heard the wrong way, a delivery read as a sound or the reverse. Phrase a delivery as a quality of the voice, such as [low, gravelly voice] or [quietly, with controlled fear], so it cannot be taken for a sound cue.

  • Experimental tags vary more from voice to voice: [strong French accent] (any accent works), [sings], [woo]. Test them before relying on them.
  • Lower Stability when tags come out too subtle; raise it when a voice overacts.
WriteWhat it does
[excited]Bright, quick, lifted energy.
[sarcastic]Dry and knowing: the words mean the opposite.
[whispers]A breathy whisper.
[shouting]A raised, projected voice.
[thoughtful]Slower, weighing each word.
[curious]An interested, questioning lift.
[annoyed]Clipped and irritated.
[appalled]Shocked disapproval.
[mischievously]Playful, up to something.
[crying]Tearful and unsteady.
[happy] [sad] [angry] [surprised]The basic emotions, each coloring the words after it.
[under his breath]Muttered, as if not meant to be heard.

Laughs, sighs, and breaths

Non-verbal tags are the sounds people make between words. Place each one where it would happen: after a line for a reaction, before it for a run-up, in the middle for a break. A laugh can build across a line in steps.

WriteWhat it does
[laughs]A full laugh.
[chuckles]A small, low laugh.
[giggles]A light, high laugh.
[starts laughing]Breaks into a laugh mid-line.
[laughs harder]A laugh already going, growing.
[wheezing]Laughing so hard the breath runs out.
[sighs]Tired, relieved, or fed up, depending on the line.
[exhales sharply]A short, hard breath out: frustration or surprise.
[inhales deeply]A big breath in before speaking.
[clears throat]A throat-clear before a line.
[gulps]A nervous swallow.
[snorts]A snort of laughter or contempt.

A laugh that builds

Okay, okay, I'll tell you. [chuckles] So the goat... [starts laughing] the goat ate the WEDDING invitations. [laughs harder] All two hundred of them. [wheezing] I'm sorry, I can't.

Why it works: The laugh grows in four steps, each where it would really happen, while the words stay intact around it.

Pauses, emphasis, and pronunciation

Eleven Speech does not read SSML, so break tags in angle brackets do nothing. Use an ellipsis for hesitation and weight, a dash for a short break, a line break between beats, or a pause tag.

Stress a word with capitals ("It was a VERY long day") and set the rhythm with question marks, exclamation marks, and ellipses. That is all the emphasis the model needs.

For a name or a term the voice gets wrong, write it in IPA with stress marks, between forward slashes and inside double quotes: The term "/ˌbaɪoʊˈkemɪstri/" covers it. Wrap only the words that need it. A plain respelling, such as "Shiv-awn" for Siobhan, also works. Results vary by voice, so listen before you rely on one.

  • Write everything as it is spoken: "January second" for 01/02, "one hundred percent" for 100%, "Doctor" for Dr., and "mage dot space slash pricing" for a web address.
  • Never put an @ in the text: Mage reads @ followed by letters as a saved handle, and a handle you have not saved stops the generation before anything is charged. Write "at" instead, as in "find us at mage space".
WriteWhat it does
...A hesitation, or weight before what comes next.
—A short break, or a line cut off mid-thought.
[short pause]A beat of silence.
[long pause]A longer, deliberate silence.

Sound effects inside the speech

A tag can also place a sound inside the read: [applause], [clapping], [door slams], [explosion], [gunshot]. Use one when the sound belongs to the moment, like a door before a line.

For a music bed or a full soundscape, make it on Eleven Music or Seed Audio and mix it under the speech in an editor.

A sound, then the voice

[applause] Thank you, thank you. [laughs] Please, sit down... [clears throat] Tonight, we're giving away something no one here has ever seen.

Why it works: The applause opens the moment, then a laugh and a throat-clear lead the speaker into the line.

Conversations between voices

One generation speaks in one voice, and a label like "Speaker 1:" or "Mara:" is read aloud. For a conversation, choose a voice for each speaker, generate each speaker's lines separately in their own voice, then join the clips in order in an audio or video editor, with short silences for the gaps.

Put a speaker's consecutive lines into one generation rather than one clip per sentence. Write interruptions across the cut: end the cut-off line with a dash, and start the reply with [jumping in] or [interrupting].

  • Keep the same voice for a speaker in every clip, and the same Stability, so the conversation sounds like one recording.
  • For an interruption, end the first clip on a dash ("So I was thinking we could—") and open the reply, in the other voice, with a tag and a dash ("[jumping in] —test the new timing?").

One speaker's turn in a two-voice scene

[nervously] So... I may have tried to fix the boiler myself. [sheepish laugh] It's making a NEW noise now.

Why it works: One voice, one turn, no speaker label: generate it on Maisie, generate the reply in another voice, and join the two.

Other languages and accents

Write the text in the language you want spoken. Every voice speaks it fluently, without carrying its own accent over, and the tags can stay in English.

To force an accent, try a tag such as [strong Italian accent]. Accent tags vary by voice, so generate a couple of takes.

Common mistakes

Avoid this, do this instead

  • Avoid: Stage directions written into the text.

    She whispered nervously: Is anyone there?

    Instead: Tags, which are performed instead of read.

    [whispers] [nervously] Is anyone there?

  • Avoid: Tags for what you see: [smiling], [walks away], [music].

    Instead: Tags for what you hear: [warmly], [chuckles], [sighs]. Make music on Eleven Music.

  • Avoid: SSML break tags.

    Instead: An ellipsis, a dash, a line break, or [short pause].

  • Avoid: Digits, symbols, and web addresses.

    Tickets are $45 at 14:30. Details at mage.space/events.

    Instead: Everything written the way it is said.

    Tickets are forty-five dollars at two thirty PM. Details at mage dot space slash events.

  • Avoid: Speaker labels in one generation: "Speaker 1: Hi. Speaker 2: Hey."

    Instead: One generation per speaker, joined in an editor.

  • Avoid: A tag that fights the voice or the meaning of the line.

    Instead: A voice that already has that delivery, and tags that match the words.

  • Avoid: A vague one-word tag that comes out as a sound effect.

    Instead: A quality of the voice: [low, gravelly voice].

  • Avoid: Changing the words while you tag.

    Instead: The script exactly as written, with tags added around it.

FAQ

Eleven Speech prompting questions

What are ElevenLabs audio tags?
Words in square brackets that direct the voice instead of being spoken: [whispers], [laughs], [sarcastic], [sighs], or a phrase like [Low, steady voice, restrained urgency]. Put one right before the words it changes, or right after them for a reaction.
How long should an Eleven Speech prompt be?
As long as the speech. Every word outside the tags is spoken, and the clip lasts as long as the text takes to say, about 2.5 words a second at a normal pace. A long script reads best in paragraphs, each opening with a direction tag.
How do I add a pause in Eleven Speech?
With an ellipsis for hesitation, a dash for a short break, a line break between beats, or a [short pause] or [long pause] tag. SSML break tags are not supported.
How do I make Eleven Speech whisper or laugh?
Put [whispers] before the words to whisper and [laughs] after the line that earns it. If a voice ignores a tag, move Stability toward Expressive or choose a voice that suits the delivery.
How do I fix the pronunciation of a name?
Write the name in IPA with stress marks, between forward slashes and in double quotes, such as "/ˌbaɪoʊˈkemɪstri/", or respell it the way it sounds. Listen to the result, since it varies by voice.
Can Eleven Speech do two voices in one clip?
No. One generation speaks in one voice, and speaker labels are read aloud. Generate each speaker separately in their own voice and join the clips in an editor.
Why did my tag come out as a sound effect?
Eleven Speech makes sound effects too, so a short tag can be heard as one. Rephrase it as a quality of the voice, such as [low, gravelly voice] instead of [gravel].

Make your first audio with ElevenLabs v4

New accounts get 300 free Gems, enough to try Eleven v4. No install, no credit card.

Try Eleven v4