Gemini 3.8 text-to-speech lets you design a voice from a description, or copy yours from 30 seconds

Google has released two new Gemini models that turn text into speech: Gemini 3.8 Flash TTS, and Gemini 3.8 Flash Lite TTS. They became generally available in the Gemini API on September 22nd. The news is that you're no longer choosing from a fixed set of voices. You can describe a new one, or copy a real one.

Slide text

01 02 03 What shipped

Gemini API release notes · Sep 22, 2026 · blog.google · Gemini 3.8 text-to-speech

Gemini 3.8 speech: design a voice, or copy yours

Generally available September 22, 2026 Gemini API and AI Studio

The two models split the work. Flash TTS is the creative one, for audiobooks, characters and scripted dialogue, in a hundred and thirty languages. Flash Lite is the fast, cheaper one, for dubbing, voice agents and read aloud features at volume. At launch prices, ten seconds of Flash audio costs about a quarter of a cent, and Lite costs less. Google's pricing page says prices for both double on January 1st.

Now the voices.

With Flash TTS, you describe the voice you want in plain words, like a high energy DJ from Melbourne, or a fire breathing dragon, and set its role, accent and character. It works across more than a hundred languages and dialects. If you'd rather not design one, there's a library of more than two thousand ready-made voices, including regional ones like Quebec French and Scots English. Voices you make can be saved and reused, so a character sounds the same next week.

You can also copy a real voice, your own or one you have the rights to, from about thirty seconds of audio. Before Gemini will create it, the owner has to record themselves agreeing, and that recording has to match the voice being copied. And every clip these models produce carries Google's SynthID watermark, which you can't hear but which marks it as generated speech. Copied voices also come with C2PA content credentials.

Once you have voices, you direct them. Each line can carry its own direction, from a calm customer service agent to a whispered suspense scene, plus laughs, sighs and little "mm hm" reactions. Two speakers can share one script and take turns naturally. And Google says a voice holds steady across hours of audio, which is what podcasts and audiobooks need.

Finally, who can use them.

Developers can use both models now, in the Gemini API and in Google AI Studio, which has a new workspace for designing and copying voices. If you don't build anything, you'll hear them in Google's own apps: Flash in Gemini Notebook, and Flash Lite in Google Vids. Gemini Enterprise customers are told it's coming soon.

If you already use Gemini's earlier speech model, the 3.1 Flash TTS preview, Google recommends Flash Lite as its replacement, and marks the old one deprecated, with no shutdown date yet. One change catches people out. The new models treat your text strictly as a script, so a direction written inline, like "say cheerfully", may be read out loud. Directions now go in a separate field, and Google's migration guide shows how.

One limit: voice remixing, where you'd take a library voice and adjust it with a prompt, is still marked coming soon. Google's announcement and the Gemini API release notes have the details. That's Gemini 3.8 speech: design a voice, or copy yours.