They exist because delivery is positional. A single "sound excited" setting applies to a whole take, but real reads change within a sentence: the aside is quieter, the punchline is faster, the sigh lands between two clauses. A tag sits at the exact point the change should happen.
The vocabulary and the syntax belong to each provider. ElevenLabs' Eleven v4 and Eleven v3 take a short instruction in square brackets, such as [curious], [whispers] or [clears throat], placed where the delivery should change. xAI's Grok TTS has two kinds: an inline tag such as [pause] or [laugh] that makes a sound at one point, and a wrapping tag such as <whisper>...</whisper> that changes the passage between its opening and closing tag. There is no shared standard, so a script marked up for one model is not portable to another.
Not every model supports them, and on a model that does not, a tag is just more text: it may be read aloud or dropped. Some models without inline tags take direction through a separate style or emotion field instead, which applies to the whole take.
In practice
- Place a tag immediately before the phrase it modifies, not at the top of the script.
- Paired tags must be closed, or the delivery change runs to the end of the take.
- Use them sparingly — a tag on every clause produces a read that lurches.
Audio tags: common questions
What are audio tags?
Short cues written into the script of an AI voice, usually in brackets, that change how the next words are said or add a sound at that point: [whispers] before a quiet line, [laughs] after a joke, [pause] before a reveal. The speech model reads the cue as direction rather than as words to say.
Which AI voices support audio tags?
Support is per model, not per app. ElevenLabs says Eleven v4, Eleven v3 and Eleven v4 Turbo support its square-bracket audio tags, and xAI documents inline and wrapping speech tags for Grok TTS. Check a model's own documentation before tagging a script for it, because each vendor has its own tag names.
The mistake to avoid
Copying a tagged script between providers. On a model that does not parse them the brackets are pronounced, and on one with a different vocabulary they are ignored.
Where you will run into it
- Add a Voiceover to a Video — Type the script. Get a narrated video back.
- Eleven v4 — ElevenLabs Eleven v4 — the most expressive ElevenLabs voice model. Direct the delivery with inline audio tags ([excited], [whispers], [laughs], [sighs], [sarcastic]), 90+ languages, and speak in your own cloned voice.
- Grok TTS — xAI Grok TTS - high-quality text-to-speech with 5 expressive voices, 21 languages, and speech tags support. Up to 15,000 characters per request.
Related terms
Text-to-speech
Text-to-speech converts written text into spoken audio using a synthetic voice you choose before generating.
Voice stability
Voice stability is the control that decides how much a synthetic voice varies its delivery — steady and predictable at one end, expressive and unpredictable at the other.
Voice design
Voice design creates a new synthetic voice from a written description — age, accent, texture, energy — instead of cloning one from a recording.
Voice cloning
Voice cloning meaning: building a reusable synthetic voice from a real sample so new scripts can be spoken in that voice later.
Speech-to-speech
Speech-to-speech takes a recording of one person talking and re-renders it in a different voice, keeping the original performance intact.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync, in your browser or on your phone.
One account for the web and iPhone.