Eleven v3 audio tags

Overview#

ElevenLabs eleven_v3 reads square-bracket cues in the text as performance direction, not as words to speak. [whispers] Something is coming is whispered. The same string on Flash, Turbo, or multilingual v2 is usually read aloud.

OnCore uses those tags in two ways:

1. Emotion tagging writes one tag from a closed set onto each English reply.
2. You write tags yourself in firstMessage, or you tell the model to emit them, when deliveryMode is "none".

This page is the tag reference. Enable emotion tagging on its own page. Voice settings for eleven_v3 (stability only) live on eleven_v3.

eleven_v3 does not support SSML. Pauses, emphasis, and emotion all go through audio tags, punctuation, and capitalisation.

Prerequisites#

• An ElevenLabs eleven_v3 voice from GET /voices. See Choosing a voice.
• Emotion tagging, if you want OnCore to write the tags: deliveryMode set to "emotion_tagging", English calls only.
• deliveryMode set to "none" if you want the model to emit tags in later replies. A tag you write into firstMessage still passes through as written.

Syntax#

A tag is lowercase or natural-language text inside square brackets, sitting inline with the words it directs.

[whispers] I never knew it could be this way.

[frustrated sigh] I was about to start over. [happy gasp] Then it clicked.

You can stack tags on one line. Match the tag to the voice. A soft-spoken voice asked to [shouts] often reads the word "shouts" instead of shouting. ElevenLabs documents that failure and recommends a voice with a wide emotional range for tag-heavy scripts.

Do not write markdown links next to a tag. [laughs](https://example.com) looks like a link to a markdown filter. Keep the form [laughs].

What emotion tagging injects#

When deliveryMode is "emotion_tagging", OnCore writes exactly one of the tags below onto the first classifiable sentence of an English reply. It does not pick from the rest of this page.

That is the full injected set. The greeting is not tagged. Later sentences in the same reply are not tagged. See Emotion tagging for the skip rules.

While that mode is on, do not also prompt the model to emit tags. You will stack two deliveries or fight the classifier.

Write tags yourself#

OnCore does not strip [tag] syntax. A tag you put in firstMessage reaches eleven_v3 as written. A tag the model emits in a later reply does too, as long as deliveryMode is "none".

Use this when the opening line needs a specific performance, or when emotion tagging is off and you want the model to direct a few moments itself.

curl -X PATCH https://core-api.heysadie.ai/assistants/0195f1e2-0000-0000-0000-000000000000 \
  -H "Authorization: ApiKey YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "firstMessage": "[warmly] Thanks for calling Joe'\''s. How can I help?",
    "settings": {
      "conversationConfig": {
        "deliveryMode": "none"
      }
    }
  }'

If you prompt the model to add tags, be specific and sparse. Use a calm baseline, stronger tags only when the moment warrants them, and no mood change mid-sentence.

A prompt fragment that works:

When you speak, you may add at most one ElevenLabs audio tag per reply,
placed at the start. Stay on [calmly] unless the caller shares good news
([cheerfully]) or a problem we caused ([apologetic]). Never invent sound
effects. Never write the tag as words.

Do not put [gunshot], [explosion], or accent jokes in a phone assistant. Those tags exist for media. They do not belong on a customer call.

ElevenLabs catalogue#

ElevenLabs publishes a non-exhaustive list. The model also accepts similar tags that describe something auditory. The lists below are the ones on ElevenLabs's prompting guide for v3, checked on 3 September 2026. Effectiveness depends on the voice and its training samples.

Voice and emotion#

ElevenLabs lists these as vocal delivery and emotional expression:

• [laughs], [laughs harder], [starts laughing], [wheezing]
• [whispers]
• [sighs], [exhales]
• [sarcastic], [curious], [excited], [crying], [snorts], [mischievously]

Their help article uses the same idea with a shorter set: [curious], [crying], [mischievously], [whispers], [shouts], [laughs], [clears throat], [sighs].

Their enhance-prompt examples add directions such as [happy], [sad], [angry], [whisper], [annoyed], [appalled], [thoughtful], [surprised], and non-verbals such as [laughing], [chuckles], [clears throat], [short pause], [long pause], [exhales sharply], [inhales deeply]. Notice [whisper] and [whispers] both appear in ElevenLabs's own pages. Try the form that matches your voice.

OnCore's injected set ([cheerfully], [calmly], [warmly], [delighted], [playfully], and the rest of the table above) is not a copy of this list. Those tags were chosen because they land on a service call. You can still write the ElevenLabs forms by hand when emotion tagging is off.

Delivery and pacing#

ElevenLabs's audio-tag guide groups volume and energy here. Examples they give: [whispers], [shouts], [softly], [rushed], [drawn out], [pause].

Punctuation also steers v3:

• Ellipses (...) add a pause and weight
• Capital letters add emphasis (VERY)
• Commas and full stops keep a natural rhythm

v3 does not support SSML <break> tags. Use an ellipsis, a [pause], or a [short pause] / [long pause].

Human reactions#

Examples ElevenLabs gives: [laughs], [clears throat], [sighs], [gasps], [gulps], [swallows].

Accents and character#

Experimental. ElevenLabs lists [strong X accent] (replace X, for example [strong French accent]), [sings], [woo], [fart], and character voices such as [pirate voice]. They warn that these are less consistent across voices. Test before you rely on one on a live line.

Sound effects#

ElevenLabs lists [gunshot], [applause], [clapping], [explosion]. These are for media and games. Leave them out of a phone assistant.

ElevenLabs also tells the enhance prompt not to invent tags for things that are not auditory ([standing], [grinning], [pacing], [music]). The same rule is safe for your own prompt.

Stability and voice choice#

stability is the only eleven_v3 voice setting OnCore sends. The three named points ElevenLabs tunes around:

For tag-heavy delivery, ElevenLabs recommends Creative or Natural. Robust will often ignore a tag you were counting on.

Pick a voice whose range matches the tags you will use. A whispering voice will not shout on command. Professional Voice Clones are, in ElevenLabs's own words, not fully optimised for v3. Prefer a designed voice or an Instant Voice Clone if tags matter.

See eleven_v3 for how to set stability on the assistant and how to override it for one call.

Pronunciation on v3#

eleven_v3 understands IPA wrapped in forward slashes in the spoken text: "/ˌbaɪoʊˈkemɪstri/". That is ElevenLabs's native form. It is separate from an OnCore dictionary pronunciation entry.

OnCore dictionaries still apply on eleven_v3 as respelling. Use a dictionary when the word should always be said a certain way. Use inline IPA when you are writing one scripted line.

What happens when a tag fails#

Limits#

• Tags are an eleven_v3 feature. Other ElevenLabs models on OnCore do not perform them.
• Emotion tagging injects only the closed set in the table above, on English replies, and never on the greeting.
• ElevenLabs's published list is not exhaustive and is voice-dependent. There is no supported API for asking OnCore which tag fired.
• Multi-speaker dialogue tags ([interrupting], [overlapping], Speaker 1:) are an ElevenLabs studio and dialogue-API feature. An OnCore assistant has one voice on the call. Do not write a two-speaker script into firstMessage.

Next steps#

• Emotion tagging: turn on the classifier that writes the closed set above.
• eleven_v3: stability, and the settings this model ignores.
• Expressive mode: model-steered delivery on Cartesia, Fish Audio, xAI, and Inworld. Not available on ElevenLabs.