AI Voice Emotion Tags: Direct Narration with [Brackets]

Sound tags: [laugh], [sigh], [breathe]

FullCastBook supports three sound tags: [laugh], [sigh] and [breathe]. Put each one exactly where the sound happens: “I told you it would rain. [sigh] Fine, take my coat.” A sound at the start of a line colors everything after it; one in the middle lands as a reaction.

Use them where a person would really make the sound, and let the text earn it. A [laugh] after a line that isn't funny sounds like a glitch, not a character. Like any direction, a sound tag is a request rather than a guarantee, so listen to every tagged line once.

Direction notes: short, specific, plain English

A delivery note describes how the line should sound, not what it means. Good notes are short and concrete, an emotion plus a manner: [hurt, trying not to show it] or [whispering, like sharing a secret]. A vague note such as [emotional] gives the voice nothing to act on, and a long one may be cut short.

In the studio, every line has a delivery field next to its speaker. When FullCastBook casts a chapter, it fills in a few words wherever the text makes the tone clear, for example a line followed by “she whispered”. Check those suggestions, clear the ones you don't want and add your own at the turning points. That is most of what bracket directions for TTS need to be.

Which voices follow tags: expressive premium vs standard

Voices differ in how they take direction. FullCastBook's standard voices run on Google's Gemini TTS and receive your note as a style instruction for the line. Sound tags such as [sigh] or [laugh] work there too: FullCastBook sends them in the angle-bracket form Google documents for Gemini TTS (<sigh>, <laugh>). The expressive premium model, ElevenLabs v3, receives a short note as an audio tag in front of the line, and only notes of 40 characters or fewer are passed on, so keep them tight. The faster premium model, ElevenLabs Turbo 2.5, reads the text without delivery notes or sound tags (the tags are left out, not read aloud): use it for even narration and switch models for heavily directed scenes.

That is also why one company's tag list doesn't transfer to every model. As of October 2026, ElevenLabs' help center says audio tags are supported by Eleven v4 and Eleven v3, so the long lists of ElevenLabs audio tags for audiobooks don't apply to its older models, and other companies document their own syntax. Even on a model that supports tags, the voice matters: ElevenLabs notes that a serious, professional voice may not respond well to playful tags.

Tags by genre: three short samples

Thriller. Narrator: The footsteps stopped outside the door. Mara: [whispering, terrified] “Don't move. [breathe] They can hear us.”

Romance. Theo: [warm, teasing] “You kept the ticket stub?” June: [laugh] “Shut up.” [sigh] “Yes.”

Horror. Narrator: [slow, hushed dread] The cellar light had been off for years. Someone had switched it on. Notice what isn't tagged: the narrator's first line in the thriller, and everything after the reveal in the horror sample. The untagged lines are what make the tagged ones land.

Overdoing it: when tags hurt narration

The most common mistake with AI voice emotion tags is tagging every line. A chapter where each sentence carries a note sounds like a performance of emotions rather than a story, and constant laughs and sighs turn into tics. Most lines read well as written; mark the turning points of a scene and leave the rest.

Keep the narrator lightly directed. Listeners spend most of the book with the narrator, and an even, warm narration makes the characters' directed lines stand out. Watch for contradictions: if the text says “she said bitterly” and the note says [cheerful], the listener hears a mistake. For voices themselves, see how to do character voices.

Regenerate one line until it lands

AI voice emotion tags rarely land perfectly on every line the first time. When a directed line misses, change one thing at a time: shorten the note, swap the emotion word, move the sound tag, or add punctuation such as a comma or an ellipsis where the pause belongs. Then regenerate only that line; the rest of the chapter stays as it is.

Each regeneration voices the line again and uses credits again, so settle your directing style on a sample chapter before you generate the book. Credits are counted by the characters voiced, so a tag adds only a few; pricing has the details. If you're building a whole cast of performances, the multi-voice text to speech page shows how the voices fit together.