ElevenLabs v3 Guide 🎙️

Eleven v3 is an emotionally rich text to speech and text to dialogue model. This guide is about how to write input text so the model performs at its best.

🧩 Mental model

• Text is the script
• Voice settings (stability, similarity) only fine tune consistency
• Emojis, punctuation, line breaks, and tags shape rhythm and emotion
• Non-speech audio and background can be hinted with bracketed descriptions
• Max input for v3 is about 3,000 characters → chunk longer content

✏️ Core writing rules

• Write like a real script or book, not a data dump
• Use normal punctuation: commas, periods, question marks, exclamation marks
• Avoid giant walls of text → use short paragraphs and clear sentences
• Keep one main idea or beat per sentence
• Avoid noisy markup that the listener should not hear (HTML, JSON, etc.)
• When you need structure (chapter, scene, section) separate with blank lines

Example (good)

The wind howled outside the cabin.
 
"We should have left earlier," she whispered.

Example (weak)

The wind howled outside the cabin she said we should have left earlier but he did not answer and then
there was a loud sound somewhere outside maybe a tree or the roof or an animal.

🎭 Emotion through words (not just settings)

Eleven v3 reads emotion mostly from the words and context. Use clear emotional cues instead of relying only on knobs.

Patterns that work well

• Emotional verbs: murmured, shouted, begged, snapped, whispered, sighed
• Adverbs and phrases: softly, through gritted teeth, barely holding back tears
• Clear intent: asking, demanding, confessing, threatening, joking
• Reactions: gasped, laughed, stammered, voice shaking

Example: neutral vs emotional

Neutral: I cannot believe you did that.
 
Upset: I cannot believe you did that, after everything we went through.
 
Devastated: I cannot believe you did that to me. After everything we went through... how could you?

🔡 Capitalization and emphasis

• Normal case → normal emphasis
• Occasional ALL CAPS on a single word can signal strong emphasis
• Do not write whole sentences in ALL CAPS
• Avoid repeating exclamation marks everywhere

Examples

Subtle: "Please, do not go."
 
Emphasis: "Please, do NOT go."
 
Too much (avoid): "PLEASE DO NOT GO!!!"

⏱️ Pauses, rhythm, pacing

• Comma , → short pause
• Period . → normal stop
• Ellipsis … → trailing off or hesitation
• New line → natural breath or beat
• Blank line → bigger change of scene or time

Example

"Wait," she said, taking a breath.
"Listen to the rain for a second... it almost feels peaceful."
 
The room fell quiet.

You can also hint at pauses with descriptive text in brackets

"I... I do not know," he said. [long pause]
"Maybe we should start over."

Timing is approximate: the model interprets it, it is not a hard timing control.

🎧 Non-speech audio and background tags

Eleven v3 supports non-speech audio behavior suggested by descriptive tags. These are written in square brackets and describe what you want to hear.

Patterns

• [laughter]
• [soft laughter]
• [sigh]
• [crowd cheering]
• [rain softly in the background]
• [door slams]

You can put them on their own line or inline.

Example

[rain softly in the background]
 
Narrator (calm): The storm had finally reached the city.
 
[crowd cheering]
 
Announcer (excited): Ladies and gentlemen, welcome to the championship match!

For subtle effects, keep the description short and natural

[sigh]
"I guess this is goodbye, then."

🗣️ Single-voice TTS vs Text to Dialogue

• Single-voice TTS

  • Great for narrators, single-character monologues, announcements
  • You can still include multiple characters but the same voice will read them

• Text to Dialogue with Eleven v3

  • Designed for multi-speaker conversations
  • Lets the model handle timing and emotional back-and-forth

If you are using Text to Dialogue, structure your script like a play.

👥 Multi-speaker dialogue structure (Text to Dialogue)

Use clear, consistent patterns so the model understands who speaks.

Good pattern

[SFX: soft rain in the background]
 
Alice (thoughtful): I never expected the city to sound this quiet.
 
Bob (light, joking): That is because you finally turned off your notifications.
 
Alice (smiling): Maybe silence is not so scary after all.

Guidelines

• One speaker per paragraph
• Put the character name at the start: Name: or Name (tone):
• Add tone hints in parentheses only when needed: (angry), (tired), (whispering)
• Avoid changing how you spell the name (do not mix “Bob” and “Bobby” for the same voice)
• Keep each line to one main idea to avoid rambling delivery

📖 Long-form narration and audiobooks

Eleven v3 is strong at long-form emotional narration, but the input should still be chunked.

Practical rules

• Keep each request below the model character limit (around 3,000 chars)
• Split by natural boundaries: scene, section, page, paragraph
• Do not break mid-sentence
• Keep typographic quotes and punctuation consistent
• Do not flood the text with markup or inline technical notes

Example chunk

Chapter 3
 
The road to the village twisted through the mountains, cutting between silent pines and crumbling stone.
For hours, they walked without saying a word. The only sound was the crunch of gravel under their boots.
 
"Do you think anyone is still there?" she finally asked.
"We will find out soon enough," he replied.

🌐 Multilingual text

Eleven v3 supports many languages.

Tips

• Write the utterance directly in the target language
• Keep one language per sentence when possible
• If you must switch, signal clearly in the text

Example

Narrator (French): La pluie tombait doucement sur les toits de la ville.
 
Narrator (English): The rain fell softly on the rooftops of the city.

For code-switching in dialogue, keep the pattern predictable

Maria: Lo siento... I did not mean to hurt you.

🎚️ Voice settings and prompting

Voice settings complement your text, they do not replace good writing.

• Stability (0.0–1.0)

  • Higher: more consistent, safer, less variation in delivery
  • Lower: more expressive and variable, but risk of sounding unstable

• Similarity boost (0.0–1.0)

  • Higher: matches the reference or base voice more closely
  • Lower: allows more freedom and variation

Starting points for expressive characters

• Stability around 0.4–0.7
• Similarity boost around 0.7–0.9

If outputs feel too chaotic

• Increase stability a bit
• Simplify the text and stage directions

If outputs feel too flat

• Lower stability slightly
• Enrich the text with stronger emotional cues and clearer verbs

🧪 Iterating on a line

A simple workflow to refine any line or scene

1. Baseline

"I need you to trust me," she said.

2. Add context

She took a deep breath, trying to keep her voice steady.
"I need you to trust me," she said.

3. Shape emotion

She took a deep breath, her voice trembling.
"I need you to trust me," she whispered.

4. Add a reaction or non-speech cue

She took a deep breath, her voice trembling.
"I need you to trust me," she whispered.
 
[silence]
 
He did not answer.

Listen, adjust one thing at a time, and keep versions that sound good.

🚫 Common anti-patterns and fixes

• Giant paragraphs without punctuation

  • Fix: split into sentences and paragraphs, add commas and periods

• Script directions you do not want spoken

  • Fix: move technical notes outside the text you send to the API

• Overusing ALL CAPS and exclamation marks

  • Fix: use strong wording and specific emotion instead

• Mixing many different languages in one sentence

  • Fix: keep language switches at clause or sentence boundaries

• Very long monologues in a single request

  • Fix: split by scene or paragraph and process in chunks

📦 Reusable prompt patterns

Single narrator, warm and calm

Narrator (warm, calm): Tonight, we will explore three small stories about finding hope in unexpected places.

Narrator with gentle humor

Narrator (light, amused): It was supposed to be a quiet evening. Then the toaster caught fire.

Two-character argument

Lena (angry): You never listen to what I am actually saying.
 
Marco (defensive): That is not fair, I am trying.
 
Lena (hurt): Trying is not the same as being here when I need you.

Game NPC line

Guard (firm): Halt. Only authorized personnel are allowed beyond this gate.

Trailer-style voice

Narrator (dramatic): In a world where the sun has not risen in seven years, one small flicker of light could change everything.

✅ Quick checklist before sending text to Eleven v3

• Is the text written like something you would actually say or read aloud?
• Are sentences clear, with proper punctuation and paragraph breaks?
• Is the main emotion obvious from the wording and context?
• For dialogue, is each speaker labeled consistently on their own line?
• Are non-speech sounds described with simple, natural tags in brackets?
• Is the text within the character limit for one request?
• Have you kept a version of prompts that produced the best results so far?