VibeVoice TTS Format Plugin — Instructions

Overview

This plugin prepares text for the VibeVoice TTS engine. It does NOT generate audio — it formats text so VibeVoice can produce the best possible speech output.

Key fact: VibeVoice does NOT support SSML or inline emotion tags. Emotion is controlled by the voice reference audio sample you provide per speaker slot.

---

format Tool

Parameters

Parameter	Type	Default	Description
text	string (req)	—	Input text to format. Plain text, "Speaker N:" format, or JSON array
mode	string	"auto"	"single", "multi", or "auto" (detect speaker prefixes)
speakers	number	1	Number of speakers (1–4). Used for voice/emotion array
voices	string[]	—	Voice names per speaker, e.g. ["Alice", "Frank"]
emotions	string[]	—	Emotion labels per speaker: neutral, happy, sad, angry, excited, calm
clean	boolean	true	Apply the text-cleaning pipeline
output	string	"script"	"plain" (no prefixes), "script" ("Speaker N:" format), "json" (structured)
emotion_map	object	see below	Maps emotion labels to speaker slots (PsiPi method), e.g. {"happy": 2, "angry": 3}

Output

---

VibeVoice Format Spec (Quick Reference)

Text Cleaning Pipeline (applied when clean: true)

Step	What
Smart quotes	"..." '...' → "..." '...'
Chinese quotes	「」『』《》 → removed
Emoji	😊 😡 → stripped
HTML tags	<br/> <p> → stripped
Non-speech annotations	[laughs] (sarcastically) → removed (emotion tags [happy] etc. are kept)
Ellipsis	… → ...
URLs	https://... → [URL]
Whitespace	Tabs → spaces, collapse multiple spaces, trim lines

Multi-Speaker Format

• Max 4 speakers (IDs 1–4)
• Lines without prefix continue the previous speaker
• Short form S1: or S 1: also works

Emotion Control

VibeVoice has no inline emotion tags. Emotion is set by the voice reference audio sample you choose:

Emotion Label	Suggested Voice Reference Type
neutral	Default reading voice
happy	Voice sample with happy delivery
sad	Voice sample with sad delivery
angry	Voice sample with angry/shouting delivery
excited	Voice sample with excited/passionate delivery
calm	Voice sample with calm/soothing delivery

PsiPi Method: Create 4 different voice reference files of the same speaker at different emotion levels, then assign each to a different Speaker slot.

Inline Emotion Annotations ([emotion])

Write [emotion] or (emotion) at the start of a line to route that line to the speaker slot mapped to the emotion. The tag is stripped from the output — VibeVoice never sees it; it only controls which speaker slot delivers the line.

With the default map this becomes:

Default emotion_map: neutral→1, calm→1, happy→2, sad→2, angry→3, excited→3.

Assign the same voice to slots 1–3 with different emotional reference clips (calm read, cheerful read, angry read) to get one speaker with dynamic emotion. Emotions missing from the map keep their original speaker slot; the tag is still removed from the output.

JSON Input Format

---

Tips

• For single-speaker text, just pass plain text with mode: "single" or mode: "auto".
• The voices and emotions arrays are documentation annotations — they don't change the text. They tell YOU which voice reference audio to select for each VibeVoice speaker slot.