ElevenLabs Eleven v3 audio-tag markup

Contract

Enforces ElevenLabs' Eleven v3 audio-tag format on the text string you send to the model. Emotion, vocal delivery, non-verbal reactions, and pauses are written as inline, lowercase, square-bracket directives ([whispers], [laughs], [short pause]) — never SSML, never parentheses, never asterisks, never capitalized stage directions. Apply whenever the task is to produce the input text for ElevenLabs' Eleven v3 (eleven_v3) voice model.

Rules

1. Bracket form. Every performance cue is a tag wrapped in square brackets with lowercase

words inside: [tag]. The brackets and the lowercase interior are the convention.

1. Inline placement. Put the tag directly at the entrance of the phrase it should affect.

A tag governs all following text until the next tag appears. Do not collect the cues into a legend, header, or trailing note.

1. A tag is a directive, not speech. The bracketed word is interpreted by the model and is

not read aloud; everything outside the brackets is the spoken text.

1. Emotion / tone tags (real, non-exhaustive):

[excited], [sad], [angry], [happily], [curious], [sarcastic], [nervous], [frustrated], [calm], [tired], [mischievously], [crying], [cheerfully], [playfully], [deadpan], [flatly], [dramatic tone].

1. Delivery-direction tags: [whispers], [shouts], [rushed], [drawn out].

1. Non-verbal reaction tags (bracketed, never spelled out as narration):

[laughs], [laughs softly], [starts laughing], [sighs], [exhales], [clears throat], [gasps], [gulps].

1. Pauses. Use [pause], [short pause], [long pause], or an ellipsis. Eleven v3 does

not support SSML <break>; pauses are bracket tags only.

1. No SSML, ever, in v3 input. Do not emit <speak>, <break>, <prosody>, <emphasis>,

<voice>, <say-as>, or <phoneme>. Eleven v3 ignores or mishandles SSML.

1. No stand-in notations. Do not use (whispering) parentheses, *laughs* asterisks, or

WHISPERING: capitalized directions — Eleven v3 reads those literally or drops them.

1. This format is for eleven_v3. The non-v3 models eleven_multilingual_v2 and

eleven_flash_v2_5 are the reverse: they take SSML <break time="1.5s" /> and do not use these audio tags. Keep the two worlds separate.

Worked examples

Base default (left) → Eleven v3 conforming form (right):

• Whisper — <prosody volume="x-soft">Come a little closer.</prosody> → [whispers] Come a little closer.
• Laugh — Ha! (laughing) Best day ever. → [laughs] Best day ever.
• Pause — Wait <break time="1.0s"/> for the drop. → Wait [short pause] for the drop.
• Excited — *excitedly* The results are in! → [excited] The results are in!
• Sigh — (he sighs) Fine, let's do it your way. → [sighs] Fine, let's do it your way.
• Sarcasm — <emphasis level="strong">Oh, brilliant.</emphasis> → [sarcastic] Oh, brilliant.
• Throat clear — Ahem. Ladies and gentlemen. → [clears throat] Ladies and gentlemen.
• Chained emotions — (curiously) What is that? (then, excited) It moved! → [curious] What is that? [excited] It moved!

Edge cases & exceptions

• Non-v3 models are the exception. If the target is eleven_multilingual_v2 or

eleven_flash_v2_5, switch to SSML: <break time="1.5s" /> (up to ~3 seconds). Those models do not interpret the audio tags. This skill's bracket format applies only to eleven_v3.

• Case. Tags are case-insensitive to the model, but write them lowercase for consistency.
• Multiple beats. Chain tags in reading order; each affects the text after it until the next.
• Sound effects are bracketed too ([applause], [gunshot]), but use them sparingly.
• Do not stack two conflicting delivery tags on the same span; pick one.

Do / Don't

• DO write [whispers] keep it down. DON'T write (whispering) keep it down.
• DO write [laughs]. DON'T write *laughs*, haha, or (laughs).
• DO write [short pause] or …. DON'T write <break time="0.5s"/>.
• DO keep the words inside the brackets lowercase. DON'T write [WHISPERS].
• DO place each tag inline before its text. DON'T list all directions in a header.
• DO reserve SSML for the non-v3 models only. DON'T mix SSML into eleven_v3 input.

Common mistakes

• Reaching for SSML (<prosody>, <break>, <emphasis>) because it is the generic,

best-known speech-markup standard — Eleven v3 does not use it.

• Encoding emotion as parenthetical or asterisk stage directions, which get spoken literally.
• Using <break time="…"/> for a pause in v3 instead of [short pause] / [long pause].
• Gathering every cue into a legend at the top instead of placing each one inline.
• Narrating reactions as prose (he laughs, she sighs) instead of [laughs] / [sighs].

Quick checklist

• ] Every cue is a [lowercase] square-bracket tag.
• ] No SSML anywhere (<speak>, <break>, <prosody>, <emphasis>, <say-as>).
• ] No (parentheses), *asterisks*, or ALL-CAPS directions.
• ] Pauses use [pause] / [short pause] / [long pause] or an ellipsis.
• ] Each tag sits inline, immediately before the text it affects.