Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS is Google's text-to-speech model for expressive, natural-sounding speech from text. It follows natural-language delivery directions for tone, pace, and emotion, renders inline tags such as laughs, sighs, and pauses, and generates two-speaker dialogue in a single request. It detects the language of the text automatically across more than 130 languages, making it suitable for narration, voice agents, podcasts, audiobooks, and localized voiceover.

Complete technical specification for integration
Ready-to-use code snippets for common workflows
Step-by-step tutorials for advanced use cases
← All GuidesPrompting
How to write text for Gemini 3.8 Flash TTS: writing for the ear, pacing with punctuation, stressing a word and keeping directions out of the transcript.
Introduction
Gemini 3.8 Flash TTS reads speech.text as a verbatim transcript: every word in it is spoken, in the language it is written in. How the words are delivered has its own field, and sounds such as a laugh or a pause go inline as tags. The text only has to hold what the listener should hear, written the way a person would say it.
Okay, so here's what changed. You open the app, tap Plan, and... that's it. Your whole week of dinners, sorted. And the shopping list? It builds ITSELF.
This guide covers the request, writing for the ear, pacing with punctuation, stressing a word, keeping directions out of the text and splitting long scripts.
The request
One field is required: the text in speech.text. Add speech.voice to choose who reads it, or the model reads with Zephyr.
import { createClient } from '@runware/sdk'
const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()
const [result] = await client.run({
model: 'google:gemini@3.8-flash-tts',
speech: {
text: 'Okay, so here\'s what changed. You open the app, tap Plan, and... that\'s it. Your whole week of dinners, sorted. And the shopping list? It builds ITSELF.',
voice: 'Sadachbia'
}
})import asyncio
import os
from runware import Runware
async def main():
async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
results = await client.run({
"model": "google:gemini@3.8-flash-tts",
"speech": {
"text": "Okay, so here's what changed. You open the app, tap Plan, and... that's it. Your whole week of dinners, sorted. And the shopping list? It builds ITSELF.",
"voice": "Sadachbia"
}
})
asyncio.run(main())curl https://api.runware.ai/v1 \
-H "Authorization: Bearer $RUNWARE_API_KEY" \
-H "Content-Type: application/json" \
-d '[
{
"taskType": "audioInference",
"taskUUID": "1a265b46-b9cb-4cf2-90d9-3b628a8363cf",
"model": "google:gemini@3.8-flash-tts",
"speech": {
"text": "Okay, so here's what changed. You open the app, tap Plan, and... that's it. Your whole week of dinners, sorted. And the shopping list? It builds ITSELF.",
"voice": "Sadachbia"
}
}
]'runware run google:gemini@3.8-flash-tts \
speech.text="Okay, so here's what changed. You open the app, tap Plan, and... that's it. Your whole week of dinners, sorted. And the shopping list? It builds ITSELF." \
speech.voice=Sadachbia{
"taskType": "audioInference",
"taskUUID": "1a265b46-b9cb-4cf2-90d9-3b628a8363cf",
"model": "google:gemini@3.8-flash-tts",
"speech": {
"text": "Okay, so here's what changed. You open the app, tap Plan, and... that's it. Your whole week of dinners, sorted. And the shopping list? It builds ITSELF.",
"voice": "Sadachbia"
}
}Response
[
{
"taskType": "audioInference",
"taskUUID": "1a265b46-b9cb-4cf2-90d9-3b628a8363cf",
"audioUUID": "316c48ab-01b3-40f1-bcc6-6dab155754e3",
"audioURL": "https://am.runware.ai/audio/os/a14d18/ws/2/ai/316c48ab-01b3-40f1-bcc6-6dab155754e3.mp3"
}
]The rest of the model builds on this request, and each part has its own guide:
- The voice:
speech.voicetakes one of 30 voice names. Choosing a voice matches them to content. - The delivery:
settings.styletakes a short direction such as "calm and unhurried". Delivery style covers how to write one. - Sounds and pauses: tags such as
<laugh>and<short pause>sit inside the text. Inline tags lists them. - Two speakers:
speech.voicesreplacesspeech.voice, and the text opens each turn with the speaker's alias. Dialogue shows the request. - The language: there is no language field. The model detects it from the text, as Languages explains.
Writing for the ear
Copy written for a page sounds read out when the model speaks it word for word. Write the way people talk: short sentences, contractions, plain words and the odd question to the listener. The small hesitations a speaker uses to think, such as "so" or "uh", go straight into the text, spelled the way they sound.
The two clips below carry the same information in the same voice.
The new budgeting feature automatically categorizes your transactions, identifies recurring subscriptions you may have forgotten about, and provides a weekly summary of your spending every Sunday.
So, the new budgeting feature. It sorts your spending for you. It finds the subscriptions you, uh, forgot you had. And every Sunday, it tells you where the money went.
Split a list into separate sentences. The written version packs three features into one sentence, and the spoken version gives each feature a sentence and a beat of its own.
Pacing with punctuation
A comma gives a short break and a period a full stop. An ellipsis (...) or a double hyphen (--) leaves a longer pause than a comma, the kind a speaker takes before the next step or while deciding what to say.
First, the onions. Low heat, ten minutes -- don't rush them. Then the garlic... just until you can smell it. That's your base.
For a pause that is part of the performance rather than the sentence, the tags <short pause> and <long pause> go anywhere in the text. Inline tags covers them.
Stressing a word
Capitalize the word you want stressed. It does what italics would do on a page and puts the listener's attention on the word that carries the meaning.
Unplug the machine BEFORE you open the back panel. The capacitor can hold a charge for several minutes after it's switched off.
Keep it to one word per sentence. When several words are capitalized, none of them stands out.
Directions and speaker labels
Because the text is a transcript, an instruction written into it can be spoken aloud with the line. A prefix such as "Say this cheerfully:" belongs in settings.style, and a speaker label such as "Maya:" is replaced by a dialogue request, where speech.voices assigns the voices.
Say this cheerfully: Welcome back! Your order is ready for pickup.The same line with the direction moved out of the text:
{
"speech": {
"text": "Welcome back! Your order is ready for pickup.",
"voice": "Laomedeia"
},
"settings": {
"style": "cheerful"
}
}Long scripts
speech.text holds up to 4,000 characters. For an audiobook chapter or a long course module, split the script at paragraph breaks and send each part as its own request with the same voice and the same style string, then join the files in order. Identical settings keep the parts consistent with each other.
Tips
-
Write what the listener should hear. Everything in
speech.textis spoken, so directions go insettings.styleand sounds go in tags. -
Stress one word per sentence. Capitals work because they are rare.
-
Keep the takes you like. There is no
seed, so the same request can come back with a different read. Generate a few and store the file you want to ship. -
Split long scripts at paragraph breaks. Send every part with the same voice and style so the parts match when you join them.