How to Use Gemini 3.8 Flash TTS: Turning Turkish Text into Speech

Gemini 3.8 Flash TTS can be used both to hand text to a plain reader and to narrate it by describing the scene. The model's announcement, the number of voices and the Flash-Lite distinction are in the Gemini 3.8 Flash TTS post. Here you'll find the order of steps for turning a Turkish text into speech without mangling it.

Write the text for speech first

A spoken script is not a blog paragraph. Spell out abbreviations: write metinden sese (text to speech) instead of TTS. Write numbers out: iki bin yirmi altı instead of 2026, and if you don't need the year, take it out of the sentence. Don't leave a proper name in the first sentence like an ordinary word. Decide yourself how the name İsmail should be read and use it consistently in the text.

Cut sentences short. A comma is a place to pause for the model. Don't pile up semicolons. Keep one idea per paragraph.

Describe the voice in a separate note

Don't mix the description with the text to be read. The note can look like this:

Voice note: calm adult narrator, Turkish, moderate pace, clear consonants, short pause at commas, no whispering, no sales excitement. Read only the script below.

Put the text to be read below it. If you write a greeting inside the note, the model may read that too. The Read only the script below line reduces this mix-up.

Flash TTS carries descriptions of emotion and pauses more fully than Flash-Lite. Start with Flash TTS for a long narration, an audiobook rehearsal or a single-character scene. If you are going to generate the same text thousands of times, Flash-Lite TTS becomes the better fit once you have locked down the description. In both, change the text without changing the description sentence. Otherwise the voice changes too.

Keep the first try short

Don't give the whole section in the first request. Four sentences are enough. Things to check: the proper name, numbers, the drop at the end of a sentence, the breath at a comma. If those are right, generate the next paragraph with the same voice note. Leave the note the same in every request. If you change the pace adjective, the narrator changes.

If two people will speak, write the owner of each line at the start of the line and ask the model for a two-speaker scene. Don't bury the speaker's name inside the line. The model may voice the name too.

Checklist

Listen with headphones, not speakers. Check whether the Turkish conjunction de is read attached to the previous word. If a foreign product name appears, write it the way you want it to sound in a Turkish sentence. The generated audio is marked with Google's SynthID watermark. Don't present it as a human voice. Cloning someone's voice is a separate consent step. Don't provide a reference recording without that step.

Save the voice note you like in a text file. Don't rely on the chat history. Setting up the same narrator next week means giving the exact same note again.

Numbers, abbreviations and proper names

In Turkish text, the place where the model drifts most is numbers. Write saat on beş otuz (fifteen thirty) instead of 15.30. Spell out the 1. list marker as birinci (first) or turn the item into a sentence. Write the percent sign as the word yüzde (percent). For currency, the abbreviation TL is sometimes read letter by letter. Saying lira sounds cleaner.

Spell out an abbreviation the first time it appears, then don't use the short form. If API repeats in every sentence, the narrator spells out letters. Build the sentence with arayüz (interface) instead. If a product name is to stay in English, don't write the pronunciation next to it. The model voices the pronunciation too. Write the name from the start the way you want it to be heard.

Pauses and emphasis

A comma is a short pause. A period closes the sentence. An ellipsis can lead the model to whisper or cut the sentence off halfway. Don't use ellipses in a narration script. Writing a word in capital letters for emphasis turns into shouting. State the emphasis in the voice note: stress the word "eşik" once. The note can carry the Turkish word in quotes. The text itself doesn't shout.

Keep the question mark in a question. Without it, the sentence is read flat and the meaning changes. Use one exclamation mark. Several in a row pull the pace toward an ad voice. If you don't want that, turn the exclamation mark into a period.

Long text

Don't give a section in a single request. Chunks of four or five sentences delay the voice from drifting. Put the same voice note at the start of each chunk. Don't write greetings between chunks. Greetings get read too and repeat when the chunks are joined.

Listen to the chunks side by side. If one is faster than another, the moderate pace from the note is missing in that request. Complete the missing note and regenerate only that chunk. You don't have to narrate the whole section again.

Before publishing the audio, listen to the proper names once more. A misread name makes the recording unusable even if the rest of the text is right. Make the fix in the text; don't try to cut and paste the audio afterwards. The cut breath is audible in the middle of the sentence.