Inworld TTS Steering Guide for LyricWinter

Inworld steering lets you write performance direction directly in the text that becomes speech. In LyricWinter, the practical version is simple: choose Inworld TTS 2 for the character, then put short bracket cues before the words that need acting direction.^[1]

LyricWinter shows that model as inworld-2. Use it when you want bracketed direction such as [whispering], [speaking slowly and sadly], or [trying not to cry] to affect the generated performance.

The warning icon on a character voice card is only a compatibility warning. It does not mean you are blocked from regenerating. It means LyricWinter noticed syntax that another model may not honor, so you can decide whether to switch the voice model backend or keep going.

Quick Start

1. Write normal speaker labels: Mira:, Theo:, Narrator:.
2. Put steering cues inside square brackets before the phrase they should affect.
3. Choose inworld-2 for that character in the voice block.
4. Use character-level regenerate when you only changed one character's voice or steering, or whole-story regenerate when you want to rebuild the full talking book.

Mira: [whispering][anxious] Wait<break time="500ms"/>now.
Theo: [cheerful] We made it.
Theo: [angry] This is BAD.
Narrator: [slow and hushed] The hallway answered with silence.

Every Way to Steer a Line

Inworld steering is natural-language direction, not a tiny fixed enum. The safest pattern is to write compact cues that a voice actor would understand, then regenerate only the clips that need another pass.^[1]

Pause Controls

Inworld pause controls use a break tag with a time value. In LyricWinter, keep the tag inside the line that owns the pause so the pause stays attached to the correct speaker.^[2]

Mira: Wait<break time="500ms"/>now.
Mira: Hold still <break time="2s" /> and breathe.
Theo: <break time="1s" /> [cheerful] We made it.

Use punctuation for ordinary rhythm and break tags for deliberate silence. Good starting values are <break time="250ms"/>, <break time="500ms"/>, <break time="1s"/>, and <break time="2s"/>.

Avoid putting a standalone break line between speakers. If the pause belongs to Mira, put it in Mira's line. If it belongs before Theo speaks, put it at the start of Theo's line.

Stacking Cues

You can stack cues when the performance needs more than one instruction. Keep the stack short and non-contradictory.

Good:
Mira: [whispering][anxious] Keep your voice down.

Good:
Narrator: [slow and hushed] The door moved again.

Too overloaded:
Mira: [whispering][shouting][sad][excited][fast][slow] I know.

A useful rule: one emotional cue plus one delivery cue is usually enough. If the result still sounds wrong, edit the line text and punctuation before adding more tags.

Copy-Ready Examples

Mira: [whispering][anxious] Wait<break time="500ms"/>now.
Mira: Hold still <break time="2s" /> and breathe.
Mira: [speaking slowly and sadly] I can still hear it.

Theo: [angry] This is BAD.
Mira: [forcing herself to stay calm] I know. Keep your voice down.
Theo: [loudly, almost shouting] You said it was locked.

Narrator: The door opened by itself.
Mira: [flat and emotionless] Great.
Mira: <break time="700ms" /> [nervous laugh] That's completely normal.

Narrator: [low and measured] The city did not sleep that night.
Narrator: [warmer] In the kitchen window, one light stayed on.
Narrator: <break time="1s" /> [quietly] And then it went out.

Voice Tags vs LyricWinter Voice Blocks

Inworld also documents voice tags, which are provider-level markup for assigning voice identity inside an Inworld request.^[3] In LyricWinter, you usually should not write raw voice tags in the story. Use the character voice blocks instead: LyricWinter detects speakers, lets you assign each character to a voice, and sends the right voice model backend during generation.^[4]

That distinction matters. Bracket steering changes how a line is performed. LyricWinter character voice blocks decide who performs the line.

Regeneration Workflow

A mismatch icon is not a stop sign. It is a prompt to review whether the selected model understands the controls you wrote. If you changed Mira from Fish Audio to Inworld TTS 2, regenerate Mira's clips to hear that character with the new backend. If you changed the whole cast or edited many lines, regenerate the full talking book.

For Inworld steering tests, change one thing at a time: first the model, then one bracket cue, then one pause value. That makes it much easier to tell whether the parser preserved your tags and whether the backend performed them.

Common Mistakes

• Using TTS 1.5 for steering: use inworld-2 for Inworld steering.
• Mixing providers: Chatterbox angle tags, Fish Audio bracket tags, Cartesia SSML, and Inworld steering are not all the same feature even when some syntax overlaps.
• Floating pause tags: put <break> inside the line that should own the silence.
• Over-tagging every sentence: steer important beats, not every clause.
• Expecting tags to fix weak dialogue: clear punctuation, sentence rhythm, and speaker context still matter.

Inworld Steering FAQ