The Voice tab reads the text in the prompt bar aloud. There are no inputs to load: you choose a voice, write the script and generate. Voice opens on Gemini 3.8 Flash.
Make a voice-over
Choose a model
Click the model card at the top of the settings panel. Models differ in voices, languages and the length of script they accept. The models table lists them.
Pick a voice and a language
Click the voice card to open the voice picker. Then set the Language select. Eleven v3 has no language select: it detects the language from your text.
Write the script
Type the words to speak in the prompt bar. The character count shows how much of the model’s limit you use. Add speech tags for pauses, sounds and delivery.
Generate
Click Generate. On models that support it, playback can start in Create before the job has finished.
A voice you make here can drive a talking character in the Avatar input mode of Video.
Choose a voice
The voice picker opens a searchable list. Use the gender filter, press the play button on a voice to hear a preview, and click a voice to select it.
- Featured shows a shorter list to start from.
- Voice library shows the full catalog, with filters for Language, Accent, Use case, Pitch and Character.
Library voices for Gemini need Gemini 3.8 Flash. Reset filters clears the search and every filter.
Speech tags
Click Tags in the prompt bar to add speech tags. They are cues the voice performs, not words it reads.
- Pauses & Breaths and Expressions insert a cue at the cursor, such as a pause, a breath or a laugh.
- Delivery sets how the words are spoken, such as a whisper or a shout. On xAI and Gemini 3.8 Flash a delivery tag wraps a phrase: select the text first, then click the tag. On the other models it goes at the cursor and sets the tone of what follows (on MiniMax, of the whole script).
The tags offered depend on the model. On xAI, Eleven v3 and Gemini 3.1 Flash you can also type your own cue in square brackets, such as [evil laugh]. Punctuation still drives pacing and intonation.
Gemini 3.8 Flash checks the script before it runs: every delivery tag must be closed, tags can only nest a few levels, and the delivery can only change a limited number of times. If a rule is broken, a message says what to fix.
Model settings
Each model adds its own controls under the model card.
| Model | Settings |
|---|---|
| Gemini | Style Instructions: describe the delivery in your own words, for example the character and mood of the speaker. |
| Eleven v3 | Stability: Creative, Natural or Robust. Output Quality under Advanced Controls. |
| MiniMax | Speed, from 0.5x to 2x. |
| xAI | Speed, from 0.7x to 1.5x. Under Advanced Controls: Output Quality (MP3 Standard, MP3 High Fidelity or WAV Studio) and Normalize Text, which reads numbers, abbreviations and symbols in spoken form. |
Models
Every voice model, with the longest script each one accepts.
GeminiGoogle- Gemini 3.8 FlashDefault
Expressive speech with precise delivery and an extended voice library
Up to 16,384 characters
- Gemini 3.1 FlashPreview
Low-latency expressive voice with live streaming and markup tags
Up to 4,000 characters
xAI- xAI Text to Speech
Expressive speech with 26 built-in voices and 20 supported languages
Up to 15,000 characters
ElevenLabs- Eleven v3Preview
Most expressive: reads inline audio tags like [laughs], [whispers], [excited]
Up to 3,000 characters
MiniMax- Speech 2.8 Turbo
Fast multilingual speech synthesis
Up to 9,999 characters
- Speech 2.8 HD
Higher-fidelity multilingual speech synthesis
Up to 9,999 characters