Google's Gemini 3.8 adds voice design and cloning to text-to-speech

Google DeepMind has released two new text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, that let users design original voices from a written description or clone a real voice from a short recording. The models are rolling out today in the Gemini API and Google AI Studio, with enterprise access coming later through Gemini Enterprise, according to the company’s blog post.

The headline feature is what Google DeepMind calls generative voice design. A user types a natural-language prompt describing a voice, such as a high-energy radio host or a fire-breathing dragon, and the model generates audio to match, spanning more than 100 languages and dialects. For those who don’t want to build a voice from scratch, the models also ship with a library of more than 2,000 ready-made voices, up from the 30 available in earlier Gemini TTS versions, and covering regional variants such as Mexican Spanish, Quebec French and Scots English.

The models can also replicate an existing voice from a 30-second audio sample, whether that’s the user’s own voice or one they have rights to use. Google DeepMind says this feature is paired with a consent check: before a voice can be cloned, the person submitting it must supply a verbal consent recording from the voice’s owner that matches the reference sample. Every clip the models produce also carries an inaudible watermark called SynthID, plus C2PA credentials, a technical standard for labeling a file’s origin, which Google DeepMind says are meant to keep AI-generated audio traceable and to protect voice performers whose voices are used as references.

Beyond generating a voice, the models let users direct how it performs. Google DeepMind says creators can write line-by-line stage directions to control pacing, emotion and dialect shifts, insert non-verbal sounds like laughs or sighs, and stage two-speaker conversations from a single script with natural turn-taking. The company says the models can hold voice quality and pacing steady across hours of continuous audio, a claim aimed at podcast and audiobook production, though it has not published independent figures to verify that consistency.

Where it ranks

Google DeepMind cites third-party benchmarks to back its quality claims. It says Gemini 3.8 Flash TTS took the top overall spot on Hume AI’s Voice Design Benchmark with a score of 71.4, and also led that benchmark’s accent-modeling category with 60.8. On Hume AI’s separate Overall Quality Index, the company says Flash TTS placed first and Flash-Lite TTS placed second. In blind listening tests on a platform called Voice Arena, Google DeepMind says the two models ranked highly in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. None of these benchmark methodologies were detailed in the announcement.

Google DeepMind named several early partners integrating the models, including Figma, HeyGen, Wondercraft and 99.co, for uses such as dubbing and voice agents. A feature called voice remixing, which would let users tweak an existing library voice’s pitch, pace or accent through prompts, is described as coming soon rather than available now. What remains to be seen is how the consent-verification system holds up against attempts to clone voices without permission, and whether the promised long-form consistency survives real-world use outside Google’s own demos.