Audio tags in prompts
Bracketed instructions inside an audio prompt: [whispers] to change a delivery, [Chorus] to mark a section. One of those is documented by its vendor and one is not.
What is an audio tag?
An audio tag is an instruction in square brackets placed inside the text of an audio prompt, rather than in a separate parameter. ElevenLabs publishes a tag vocabulary for its v3 models that changes how a line is delivered and can insert non-speech sounds. Suno's [Verse] and [Chorus] structure tags are widely used and widely documented by third parties, but we could not find an official list published by Suno, so they are recorded here as convention rather than specification.
Why they matter
- A tag is the only way to direct a performance in a model that takes no separate style parameter. Without it a text-to-speech model reads the line neutrally, and asking for emotion in prose usually gets the words "say this excitedly" spoken aloud.
- Tags are also the most model-specific thing in prompt craft. A vocabulary documented for one model family is not portable, and a tag a model does not recognise is simply text it may read out.
- They are probabilistic, not commands, in both products. Both vendors' own framing is that a tag steers a generation rather than guaranteeing it.
ElevenLabs v3 audio tags
Eleven v3 text-to-speech, per ElevenLabs' own documentation
Syntax: [tag] placed inline in the text, at the point where it should take effect
Voice and delivery
- [laughs]
- [laughs harder]
- [starts laughing]
- [wheezing]
- [whispers]
- [sighs]
- [exhales]
- [sarcastic]
- [curious]
- [excited]
- [crying]
- [snorts]
- [mischievously]
Sound effects
- [gunshot]
- [applause]
- [clapping]
- [explosion]
- [swallows]
- [gulps]
Unique and special
- [strong X accent]
- [sings]
- [woo]
- [fart]
How they behave
| Rule | What it means in practice |
|---|---|
| The voice you choose limits what a tag can do | ElevenLabs states that the voice and its training samples affect tag effectiveness, and gives the example that a whispering voice will not suddenly shout because you added a shout tag. Tag choice follows voice choice, not the other way round. |
| Some tags are less consistent than others | The documentation describes experimental tags as less consistent across voices and advises testing thoroughly before production use. Treat an unusual tag as something to audition, not something to schedule around. |
| Punctuation does work that tags cannot | Ellipses add pauses and weight, and capitalisation increases emphasis. These operate alongside tags, so a line often needs fewer tags than it first appears to. |
| One or two per line, at the point of effect | The documented examples place one or two tags per phrase rather than stacking them. A tag applies from where it sits, so its position in the line is part of the instruction. |
Suno structure tags
Suno lyric input. Widely used, and not published by Suno as far as we could verify.
Syntax: [Section] on its own line, above the lyrics belonging to that section
Song sections
- [Intro]
- [Verse]
- [Verse 1]
- [Verse 2]
- [Pre-Chorus]
- [Chorus]
- [Hook]
- [Bridge]
- [Outro]
- [End]
Instrumental and dynamics
- [Instrumental]
- [Instrumental Break]
- [Break]
- [Interlude]
- [Drop]
- [Build-Up]
- [Breakdown]
How they behave
| Rule | What it means in practice |
|---|---|
| They are hints, and the model may ignore them | Every account of these tags, including the ones that document them most thoroughly, describes them as probabilistic. Suno follows them much of the time and sometimes does not, so a structure tag is not a guarantee of an arrangement. |
| A section tag goes on its own line | Placed above the lyrics for that section rather than inline, and repeated wherever the section recurs, so a chorus that comes back three times carries the tag three times. |
| This vocabulary is not verified against a vendor document | Treat the exact spellings as convention. If a tag does nothing, the likeliest explanation is that it was never a documented tag, not that the model failed. |
The advice that contradicts itself, and how to tell which applies
Some prompt guidance says to use no emotion cues, parentheticals or stage directions in a spoken line.
Where that holds
That is correct for a pipeline where narration is plain text and the voice is selected by id, which is how several production systems are built. In that setting a bracketed cue is read aloud or ignored.
Where it does not
It is wrong for a model that documents a tag vocabulary. On Eleven v3, bracketed tags are the documented control surface for delivery, so removing them removes the only handle you have.
How to tell: Check whether the model documents tags. If it does, tag. If it does not, write the line plainly and carry the performance direction outside the prompt.
Common questions
What is an audio tag?
An audio tag is an instruction in square brackets placed inside the text of an audio prompt, rather than in a separate parameter. ElevenLabs publishes a tag vocabulary for its v3 models that changes how a line is delivered and can insert non-speech sounds. Suno's [Verse] and [Chorus] structure tags are widely used and widely documented by third parties, but we could not find an official list published by Suno, so they are recorded here as convention rather than specification.
Does ElevenLabs publish a list of audio tags?
Yes. ElevenLabs documents a tag vocabulary for Eleven v3, grouped into voice and delivery, sound effects, and a set it calls unique and special. The list on this page is quoted from that documentation, read on 2026-08-17.
Does Suno publish an official list of metatags?
Not that we could find. Structure tags such as [Verse] and [Chorus] are used and documented extensively by third parties, but we could not locate an official list in Suno's own help centre, so they are recorded here as community convention rather than specification.
Are audio tags reliable?
No, and neither vendor claims they are. ElevenLabs describes some tags as less consistent across voices and advises testing before production use, and every account of Suno's structure tags describes them as hints the model can ignore.
Should a spoken line contain emotion cues or not?
It depends on the model, which is why the advice conflicts. That is correct for a pipeline where narration is plain text and the voice is selected by id, which is how several production systems are built. In that setting a bracketed cue is read aloud or ignored. It is wrong for a model that documents a tag vocabulary. On Eleven v3, bracketed tags are the documented control surface for delivery, so removing them removes the only handle you have.
Not covered here
- Udio: no tag documentation was retrieved, and its export situation is separately unresolved — see the model page.
- ID3 and container metadata, which is a different meaning of the word tag and not prompt craft.
- Any tag vocabulary for music models other than Suno, which we have not checked.
Sources
-
Prompting Eleven v3: best practices
The tag vocabulary below is quoted from this page. ElevenLabs groups the tags itself; the grouping is theirs, not ours.
-
No official metatag list found
Checked Suno's help centre and could not find a published list of bracket tags. The structure vocabulary below is therefore recorded as community convention. If Suno publishes one, this section should be replaced with it and the convention label removed.