Direct the voice with audio tags
How bracketed audio tags, punctuation and voice choice shape delivery in Eleven v3 and v4.
Eleven v3 introduced audio tags, and Eleven v4 continues them: short instructions in square brackets inside your text that change how the next line is delivered. For short-form narration they are the difference between a flat read and a voice that actually performs.
What tags do
Tags fall into roughly three groups. Emotion and delivery tags such as [whispers], [sighs], [excited], [curious] or [sarcastic] steer the performance. Reaction tags such as [laughs] add a human sound. Sound-effect tags like [applause] create a noise instead of speech. The docs also list experimental ones, like an accent tag or [sings], that are less consistent between voices.
You can combine tags for layered delivery, but describe a voice quality clearly. A vague tag can be read as a sound effect instead of a delivery instruction.
Match tags to the voice
The voice you pick matters more than any tag. Delivery that already exists in a voice's training audio is easiest to reproduce; a calm narrator asked to [shout] may sound odd or ignore it. Eleven v4 can follow tags a voice was never trained on, but the docs say results vary, so test your own voice with your own script.
Punctuation still counts
Ellipses add a pause and weight, capital letters add emphasis, and normal punctuation sets the rhythm. For deliberate silences, v4 supports [pause] for a breath-length beat and [long pause] before a reveal. SSML break tags do not work in v3 or v4.
Watch for spoken directions
If you describe emotion in plain words instead of a tag, the model may read that description aloud. Either use a proper tag or trim the audio afterwards.
Workflow that works
Write the script first, then add tags only where the emotion changes; tagging every sentence tends to sound forced. Generate several takes of the important lines, such as the hook, and keep the best one. Creator tutorials often speed this up by asking an LLM to place the tags, then reviewing the result by ear before generating.