jordangaston ElevenLabs TTS Prompting
A method for turning plain text into a script that ElevenLabs Eleven v3 performs the way you intend. v3 reads structure — tags, punctuation, and capitalization steer emotion and pacing. This skill catalogs those controls and gives you a repeatable workflow.
Default to Eleven v3. It is the only model that reads audio tags. The v2 fallbacks (SSML
<break>and<phoneme>) appear at the end for older models only.
The workflow
Given plain text and a target delivery, produce a tagged script:
- Pick the voice character first. Tags amplify a voice; they do not override it. A meditative voice will not convincingly shout; a hyped voice will not whisper. Match the script's emotional range to a voice that already sits in that range.
- Set stability for the intended expressiveness (see table below).
- Place audio tags at the moments emotion or sound shifts — before the clause they color, not after.
- Shape pacing with punctuation using the cheat sheet — ellipses, commas, dashes, capitals.
- Fix any hard pronunciations with IPA in slashes.
- Read it back. Cut tags that fight the voice or stack too densely. Over-tagging destabilizes v3.
Deliver the script in a code block so the user can paste it straight into ElevenLabs.
Punctuation cheat sheet
Punctuation is the primary pacing control in v3. This maps each mark to its effect:
| Write this | Effect |
|---|---|
... (ellipsis) |
Long pause; adds weight and hesitation |
, (comma) |
Short pause — a breath |
? |
Raises the intonation |
CAPITALS |
Shout / excitement / emphasis |
- (dash) |
Break; a clipped stop |
Example: "It was a VERY long day [sighs] ... nobody listens anymore."
Standard punctuation gives natural rhythm — write in complete, well-punctuated sentences. Capitalize whole words only where you want force behind them.
Audio tags
Tags go in square brackets inline. Place a tag immediately before the words it should affect.
Voice and emotion — control vocal delivery:
[laughs], [laughs harder], [starts laughing], [wheezing], [whispers], [sighs], [exhales], [sarcastic], [curious], [excited], [crying], [snorts], [mischievously]
Sound effects — add environmental sound:
[gunshot], [applause], [clapping], [explosion], [swallows], [gulps]
Experimental — powerful but inconsistent across voices; test before relying on them:
[strong X accent] (replace X, e.g. [strong French accent]), [sings], [woo], [fart]
Tag rules
- Match the tag to the voice. Some tags land on some voices and not others. Do not expect
[shout]to make a soft voice roar. - Give the tag emotional context. Tags work best when the surrounding text already leans that way —
[sarcastic] Oh, great. Another meeting.beats a lone[sarcastic]on neutral words. - Do not stack tags densely. Too many tags in one generation destabilizes the output. One clear tag per emotional beat.
- Experimental tags are a gamble. Reserve
[sings], accents, and novelty tags for drafts you will audition.
Stability settings
The stability slider controls how closely the output tracks the reference voice.
| Setting | Behavior | Use when |
|---|---|---|
| Creative | Most expressive; more prone to hallucination | You want tags and emotion to land hard and will re-roll misfires |
| Natural | Closest to the reference recording; balanced and neutral | Default for expressive narration with reliable results |
| Robust | Highly stable; least responsive to prompts | You need consistency across long or batched generations |
For tag-driven expressiveness, choose Creative or Natural — Robust mutes directional cues. Use Instant Voice Clones or designed voices; Professional Voice Clones are not yet optimized for v3.
Pauses
v3 has no SSML <break> support. Control pauses with the punctuation cheat sheet — ellipses for long pauses, commas for breaths, dashes for clipped stops. This is deliberate: text structure drives pacing in v3.
Pronunciation
Wrap a phonetic spelling in forward slashes to force pronunciation. v3's IPA support reaches 80–90% consistency. Include stress markers — primary ˈ and secondary ˌ.
The term "/ˌbaɪoʊˈkɛmɪstri/" refers to the study of chemical processes.
Multi-speaker dialogue
Assign a distinct voice from the Voice Library to each speaker, then label lines by speaker. Keep each speaker's tags in their own voice's range.
Speaker 1: [excited] You will not believe what just happened!
Speaker 2: [curious] Go on...
Speaker 1: We got the grant. [laughs] We actually got it!
Text normalization
v3 reads symbols and abbreviations literally, so spell out anything you want pronounced a specific way:
- Abbreviations:
Dr.→Doctor,Ave.→Avenue - Alphanumerics:
Ctrl + Z→control z,100km→one hundred kilometers - Phone numbers, currencies, dates, and URLs: expand to words, respecting locale.
Do / Don't
Do
- Choose a voice whose natural range already covers the emotion.
- Write complete, well-punctuated sentences; let punctuation pace the read.
- Place one tag per emotional beat, right before the words it colors.
- Capitalize whole words for genuine emphasis, sparingly.
- Spell out symbols, numbers, and abbreviations you care about.
Don't
- Stack multiple tags on one line — it destabilizes v3.
- Leave stage-direction prose in the script (e.g. "she said angrily") — it gets read aloud. Encode intent as tags or punctuation instead.
- Fight the voice: no
[shout]on a whisper voice. - Use SSML
<break>or<phoneme>with v3 — they do nothing. - Lean on experimental tags in a final script without auditioning them.
Worked example
Input: "I can't believe you did that. It's the best surprise ever. Thank you so much."
Tagged for a warm, expressive voice on Natural stability:
[gasps] I ... I can't believe you did that. [laughs] It's the BEST surprise ever.
[crying] Thank you ... thank you so much.
The ellipses stage the emotional catch, [gasps] and [crying] bracket the arc, and BEST puts weight on the peak — without a single line of stage direction for the model to read aloud.
v2 fallback (older models only)
Use these only when a task pins an older model. They do nothing in v3.
- Pauses:
<break time="1.5s" />up to 3 seconds. Too many break tags in one generation cause instability. - Pronunciation:
<phoneme alphabet="cmu-arpabet" ph="M AE1 D IH0 S AH0 N">Madison</phoneme>. CMU Arpabet is more reliable than IPA for v2; the<phoneme>tag is v2.5 Flash only.