A line can need relief without laughter, urgency without shouting, or a pause without an audible sigh. Eleven v3 audio tags are useful when you can describe that difference and hear whether the model followed it.
Download the director's script and take log · Test the plain narration in Voice Lab
Editorial guide checked September 13, 2026. Examples and worksheets are original planning material; the hero is an AI-generated illustration, not a model benchmark.
Tags are performance directions, not guaranteed effects
ElevenLabs describes audio tags as bracketed instructions that influence delivery or vocal sounds. Its audio-tags introduction shows emotional and nonverbal directions. The idea is straightforward: place a relevant audible instruction near the passage it should affect.
The difficult part is choosing a direction that matches the scene. “Excited” can mean warm anticipation, breathless surprise, or loud celebration. A short bracketed label does not remove that ambiguity. The surrounding words, selected voice, and settings all affect what the listener hears.
Start by writing what the audience should understand. Then decide what the performance should add. In our fictional scene, a museum guide discovers a hidden door and invites a visitor to look closer. The voice needs curiosity and a brief quieter moment, not a sequence of theatrical sound effects.
Write the clean script before adding direction
This is the baseline. It establishes the words, logical pauses, and intended meaning without special markup. Generate or read it aloud once before adding tags. If the passage is awkward in plain speech, emotional directions are unlikely to fix the writing.
Mark the one phrase where the delivery should change: “There's a handle behind it.” That is the discovery. The final instruction should return to a calm, practical tone. Keeping those decisions explicit prevents every sentence from acquiring a different mood simply because the model can produce one.
Use the baseline as the source of truth during review. A take that adds a new sentence, changes “don't,” or replaces the curator with another person has changed the script. Do not let a compelling performance distract you from that error.
Add one direction at the point of the change
These are experimental performance directions, not measured sample results. The last line may retain more of the whisper than you intended. If so, generate the calm ending as a separate segment or use a clearly supported direction that restores the desired delivery. Listen to the join before deciding that segmentation improved the scene.
Do not add a sigh merely to create a pause. A sigh communicates an emotional or physical event. If the character should simply wait, use the model's supported pacing controls or edit the silence after generation. An unintended sigh can change the meaning from curiosity to disappointment.
Likewise, laughter should have a reason. A cheerful instructional voice can sound friendly without laughing after every sentence. Use the minimum number of audible events needed to support the scene.
Choose a voice that can plausibly perform the brief
The ElevenLabs best-practices guide identifies voice selection as a major factor in v3 performance. It describes Creative, Natural, and Robust stability behavior, with a tradeoff between expression and consistency. It also states that v3 does not support SSML break tags.
For the museum example, begin with a voice already suited to measured narration. Asking a highly energetic announcer to become intimate and restrained may create more work than selecting a closer starting voice. Evaluate the untagged read before investing in elaborate direction.
Compare a small set of takes at the same settings. If the voice is suitable but the tags are ignored, examine the model's current stability options. If the words become unreliable when expression increases, return to a more stable setting and simplify the direction. Save the actual settings rather than describing a preset only as “more emotional.”
Diagnose spoken tags and ignored tags differently
| Symptom | First check | Useful next test |
|---|---|---|
| The voice says “curious” aloud | Is the selected model really v3, and does the interface support its tags? | Try one documented tag on a short sentence in a compatible interface |
| The words are correct but the direction is ignored | Does the chosen voice and stability setting suit the requested delivery? | Keep the sentence fixed and compare a closer voice or a different supported setting |
| The voice adds extra sounds or words | Is the prompt overloaded or the performance too unconstrained? | Remove all but one direction and compare with the clean baseline |
| The emotion changes at the wrong place | Where is the tag placed relative to the intended phrase? | Move it next to that phrase or split the passage into deliberate segments |
A model mismatch is not a prompt-quality problem. Rewriting brackets repeatedly will not make an unrelated speech engine implement v3 behavior. Conversely, a compatible model can still produce an unsuccessful take; support for the syntax does not guarantee a particular performance.
Keep a tiny diagnostic passage on hand, such as one plain sentence followed by the same sentence with a single tag. It makes it easier to tell whether the problem belongs to the interface, the voice, or a complicated script.
Treat exact timing as an editing requirement
If the narration must fit a ten-second shot, establish a usable read before chasing exact duration. A natural line that is slightly long may be easier to edit than a rushed line that technically fits. Shorten the writing when the message allows it.
Do not assume that an SSML pause instruction copied from an older model will work in v3. Use supported directions and punctuation for the performance, then adjust silence in an audio editor when a precise gap is necessary. A pause before a reveal and a pause between two unrelated segments have different dramatic jobs.
When joining separately generated lines, compare room tone, loudness, pacing, and vocal intensity. Leave enough clean material around the join to make a natural edit. If the character sounds suddenly closer to the microphone, the transition may be more distracting than the original imperfect pause.
Keep a director's take log
For each take, record the plain script version, marked script version, voice, model, settings, and outcome. Separate “words correct” from “performance appropriate.” A take should pass the first requirement before you compare artistic preferences.
Use three listening questions: What emotion did I actually hear? Where did it begin and end? Did it help the intended meaning? Ask a second listener when a direction is subtle. If both people hear sarcasm in a line intended as reassurance, the tag label does not make the performance reassuring.
Keep rejected takes in the log without presenting them as proof that the model always fails. A small creative test describes the behavior of those settings on that script. It does not establish a universal success rate across voices and languages.
Use QuestStudio for the clean narration stage
QuestStudio currently lists ElevenLabs Turbo in Voice Lab, not Eleven v3. Use the plain museum script there to evaluate wording, pronunciation, and basic pacing with a supported model. Use an explicitly compatible Eleven v3 interface for the bracketed version.
That separation prevents a confusing comparison: v3-specific tags read aloud by another model tell you little about either model's narration quality. Keep the model name attached to every output.
The downloadable director's sheet includes the baseline, marked example, and listening questions. Once a take passes, save the approved words and timing with the audio. The reusable asset is a directed performance that says the right thing, not a long list of tags that happened to produce an interesting sound.
Frequently asked questions
What are Eleven v3 audio tags?
They are bracketed directions used to influence vocal delivery or sounds, such as [curious], [whispers], or [sighs]. Their effectiveness depends on the voice and context.
Why is my voice reading the tag aloud?
First confirm that the selected model and interface support Eleven v3 tags. Then simplify the direction and test it on a short passage with a compatible voice.
Can I use an SSML break tag in Eleven v3?
ElevenLabs says v3 does not support SSML break tags. Use its supported audio directions, punctuation, and text structure, then edit exact timing if needed.
Can I use these tags in QuestStudio's ElevenLabs Turbo?
Do not assume compatibility. Turbo is a different model; use a clean narration script there and test v3-specific markup in an explicitly compatible interface.