Turn Scripts into Directed Performances with Gemini 3.1 Flash TTS
Last verified: November 3, 2026
Gemini 3.1 Flash TTS is Google's text-to-speech model that takes its stage directions from the script itself. You type bracketed audio tags such as [whispers] or [laughs] straight into the text, and they steer vocal style, pace and delivery, so one script carries a whole performance . Google launched it on April 15, 2026 as the speech-generation half of Gemini 3.1 Flash Audio. That group also includes Gemini 3.1 Flash Live and is based on Gemini 3 Pro . Text goes in and audio comes out , so it works as a dedicated engine for narration and spoken app content.
What it does best is keep the direction right next to the words. A plain-language brief sets the character and the scene, and inline tags change the delivery exactly where they sit in a line. Audio tags arrived with this model, along with overall gains in naturalness, controllability and multilinguality . It supports more than 70 languages . Here, the brief, the tags and the transcript all go into one Text box.
Explore Gemini Omni's Models
Speech Generation Specs at a Glance
What goes in, what comes out, and the controls that shape each take.
Input and output
Text in, spoken audio out
Script length per run
Up to 40,000 characters
Preset voices
30 named voices, with Kore selected by default
Language setting
Auto-detect or 87 language-and-region options
Inline delivery cues
200+ audio tags, including [whispers] and [laughs]
Audio files per run
1
Six Checks Before You Press Generate
Catch the slips that ruin a take: directions read aloud, mismatched voices, misdetected languages and long-script drift.
Mark where the spoken words start
When direction and dialogue share one string, an ambiguous prompt can make 3.1 Flash TTS read your stage directions aloud . Open with a line asking for speech. Keep your notes under labels like Audio profile and Director's notes, and start the words to be spoken after a clear Transcript: label.
Cast a voice that suits the persona
Pick the Voice whose natural sound already fits the character in your brief. When the brief and the preset agree, the model has one consistent target. Try two or three presets on a single short line before you commit a long script to one of them.
Put each audio tag right before the words it changes
Tags work inside the text, so where you place one is the instruction. Put [whispers] directly in front of the phrase you want whispered. Save tags for real changes in delivery: if every sentence has one, the listener hears no contrast.
Set Language when Auto might guess wrong
Auto works out the language from the text, which suits long scripts in a single language. For one-line clips, scripts full of foreign names, or a specific regional sound such as English (UK) or Portuguese (Brazil), choose the language-and-region option yourself.
Read Enhance Prompt suggestions word for word
Enhance Prompt can expand a rough idea into a fuller prompt. Before you accept a suggestion, check that it hasn't reworded lines that must be spoken exactly as written, such as product names, quotes or legal copy.
Split chapter-length scripts into sections
The Text box accepts up to 40,000 characters, but when a brief and a transcript share the same text, the voice can slowly drift across long passages . For long narration, generate one section per run with the same Voice and Language, and listen to the end of each take before moving on.
Gemini 3.1 Flash TTS or OpenAI TTS: Pick the Engine for Your Script
Each row shows what Google and OpenAI document about their own speech models: what goes in and comes out, how many preset voices you get, language reach, how you direct the delivery and what each model is built on. Use it to match an engine to the voice work in front of you.
| Feature/Spec | Gemini 3.1 Flash TTS | OpenAI TTS |
|---|---|---|
| Input and output | Text in, audio out | Text in, spoken audio out |
| Preset voices to cast from | 30 prebuilt voices | 13 built-in voices |
| Language reach | 70+ languages | Whisper-supported languages; voices optimized for English |
| How delivery is directed | Natural-language style prompts plus 200+ inline audio tags | gpt-4o-mini-tts: a written instructions prompt controls aspects of the speech |
| What the model builds on | Part of Gemini 3.1 Flash Audio, which is based on Gemini 3 Pro | gpt-4o-mini-tts: built on the GPT-4o and GPT-4o-mini architectures |
Match Each Engine to Your Voice Work
Direction inside the script vs. a separate brief
Gemini 3.1 Flash TTS reads its direction from the same text it performs: the brief and the bracketed tags sit in one string with the words. OpenAI's gpt-4o-mini-tts takes its direction as a separate instructions prompt sent next to the input text . Each approach asks something different of the writer. With the Gemini model, notes and spoken words share one field, so the script has to make clear which is which. With gpt-4o-mini-tts, the brief lives in a field of its own.
In practice, write the Text box in a fixed order: brief first, then the transcript after a plain label. That way your performance notes are never mistaken for lines.
Casting range and language reach
Roster size matters most when a script has several characters, or when a brand wants a voice that doesn't sound like everyone else's. Gemini 3.1 Flash TTS has more than twice as many presets as OpenAI TTS , and the Voice menu here lists 30 of them.
On language, the Gemini model covers more than 70 languages. OpenAI's TTS voices follow Whisper's language support but are optimized for English. For scripts in Hindi, Portuguese or Swahili, the Language menu here lets you set an exact language-and-region option instead of relying on detection.
Pick the Engine That Fits Your Cast
Use this quick guidance to pick the best option for your workflow.
How to choose: Pick Gemini 3.1 Flash TTS when casting range matters: its 30 prebuilt voices give you more to choose from than the 13 built into OpenAI TTS. For the core job, the two are the same: both turn text into speech, and both Gemini 3.1 Flash TTS and gpt-4o-mini-tts take natural-language direction on how the words should sound.
From Script to Spoken Audio in Four Steps
Four steps take you from a blank Text box to one finished audio file.
Write your script in the Text box
Type or paste the words you want spoken. If you want a particular character, add a short brief at the top, and put bracketed cues in the transcript wherever the delivery should change.
Polish a rough idea with Enhance Prompt
Only have a one-line idea? Click Enhance Prompt under the box to see fuller prompt suggestions. Your text stays as it is until you choose one.
Choose a Voice and a Language
Select one of the 30 presets in Voice (Kore is preselected). Then leave Language on Auto, or pick a specific language and region when you need a set accent.
Click Generate and listen back
Each run returns one audio file. Play it, adjust the tags or the brief wherever the delivery missed, and generate again.