Prompt ElevenLabs for text-to-speech

Last updated: October 12, 2026

ElevenLabs voices a line in the voice you choose and takes its direction from the text itself: square-bracket audio tags for the delivery, punctuation for the pacing. Pick it for character barks, dialogue and narration where the performance matters as much as the words.

Let the agent write it

You don't have to write any of this yourself. The Layer agent knows how to prompt each model it runs, from the same guidance this page is built on. Name the model, the voice and the moment, for example: Voice this line with ElevenLabs TTS V4 in the Rook voice: he whispers to his cellmate, pauses, then turns determined. "Keep your voice down. The warden's patrol passes this cell at midnight, and I mean to be gone before then." The agent adds the tags and punctuation, shows you its plan, and runs it when you approve. Use the rest of this guide when you want to write or adjust the line yourself, and see Write a brief for the agent for how to brief it.

When to use it

Use ElevenLabs when a line needs acting: a whisper that turns into a threat, a laugh mid-sentence, a sigh before a confession. It suits character barks, cutscene dialogue, tutorial narration and trailer voice-over, in a voice you keep across a whole cast.

Every run needs a voice: a voice reference set in your workspace. Keep one voice per character and reuse it on every line, since a voice that changes between lines is more noticeable than one that is slightly off.

Versions on Layer

ElevenLabs TTS V4

The version this guide's examples were made with. ElevenLabs says it follows audio tags more closely than v3, including deliveries the voice was not recorded doing. It adds Similarity boost, a Seed for repeatable takes, and pronunciation written in IPA.

ElevenLabs TTS V3

The same audio tags and punctuation, with a Stability control but no seed or similarity setting. The voice you pick matters more here: a tag lands when the delivery is already in the voice's range.

How to write the prompt

The prompt is the script. Put the delivery in square brackets right before the words it changes, and let punctuation set the pace:

[whispers] Keep your voice down... The warden's patrol passes this cell at midnight. [pause] [quietly determined] And I mean to be GONE before then.

A tag carries forward until the next one, so the whisper holds for the first two sentences, the [pause] opens a gap, and [quietly determined] changes the read for the last sentence. The ellipsis adds a beat, and the capitals lean on "gone". That prompt is the one in the clip above.

What works

Ellipses for pauses, capitals for stress

Punctuation is how you pace a line. We voiced the same line twice on V4 with the same voice and seed, once with ordinary punctuation and once with ellipses: "Keep your voice down... The warden's patrol passes this cell at midnight... and I mean to be gone before then." The ellipsis take ran nearly a second longer, with a clear half-second gap after "midnight" where the plain take ran straight on. Capitals add emphasis: "I want it right NOW."

Tags right before the words

Write the direction in square brackets where the delivery should change: [whispers], [shouts], [sarcastic], [excited], [sighs], [laughs]. In our test on the same voice and seed, [whispers] and [shouts] came out quieter and louder than the untagged line, in that order. We judged that by measuring loudness, not by listening, so take the delivery itself from ElevenLabs' own documentation linked below.

Directions in your own words

Tags are free text. Brief the voice the way you would an actor: [quietly determined], [low, gravelly voice], [out of breath after running up the stairs]. Combine two directions with a comma inside one bracket, [whispering, playful], rather than stacking vague ones.

Pauses where you want them

Write [pause] for a clear gap or [long pause] for a held silence. In our stealth line, the [pause] after "midnight" left a longer gap than the ellipsis alone had at the same place.

A tag only at each shift

The delivery carries forward, so you don't need a tag on every sentence. Add the next one only where the performance should change.

Numbers and names written out

Write numbers, dates and abbreviations the way they should be said: "one hundred gold", "Doctor Vance". On V4, give an invented name a fixed pronunciation in IPA between forward slashes, with stress marks, so the whole cast says it the same way; results can still vary by voice.

What to avoid

Tags the voice can't perform

A tag lands when the delivery is within the voice's range. A calm narrator voice will not shout convincingly, and a bright, energetic one will not whisper well. For a tag-heavy script, pick a voice with a wide emotional range.

Round or curly brackets

Only square brackets are read as direction. A tag in parentheses or curly braces gets spoken as words.

Short, ambiguous tags

A one-word tag can be taken as a sound effect rather than a delivery. Describe the delivery: [low, gravelly voice] rather than [gravel].

SSML break tags

These versions don't read SSML pause tags. Use punctuation, [pause] or [long pause] instead.

Settings that matter

Reference set holds the voice; a run needs one. Stability sets how steady the voice is: lower gives a broader emotional range and more variation between takes, higher is consistent but can turn monotonous. Similarity boost (V4) sets how closely the output keeps to the original voice; on a voice cloned from noisy audio, a high value brings the noise along. Text normalization spells out numbers, dates and abbreviations before they are spoken. Speech language sets the language the voice speaks in. Seed (V4) repeats a take when you change one word.

Prompts to start from

Stealth whisper with a turn

[whispers] Keep your voice down... The warden's patrol passes this cell at midnight. [pause] [quietly determined] And I mean to be GONE before then.

One delivery for the whole line

[whispers] Keep your voice down. The warden's patrol passes this cell at midnight, and I mean to be gone before then.

Use it in Layer

  1. In a session, open Create speech from the side panel.

  2. Click the model name at the top of the panel and choose ElevenLabs TTS V4.

  3. In the prompt box, write the line, with tags where the delivery changes and punctuation for the pace.

  4. Click Reference set and choose the character's voice.

  5. Click Create speech. The cost shows on the button before you click.

  6. When the take lands on the board, click View to play it. Change one tag or one punctuation mark at a time and run it again.

Learn more

Related