Google Releases Gemini 3.8 TTS, Letting You Design a Voice From Scratch

Google DeepMind unveiled two new text-to-speech (TTS) models on September 23 local time: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. TTS is the technology that reads written text aloud in a human-sounding voice. Both models became available to developers through the Gemini API and Google AI Studio the same day.

Google says the release moves voice generation away from a fixed menu of preset voices and turns it into something closer to a creative studio. Flash TTS is aimed at character design and fine-grained performance direction, while Flash-Lite TTS targets cost-sensitive work such as bulk dubbing and voice agents.

Building a voice from a prompt

The biggest change is that you can now design a voice yourself. Instead of choosing from the previous set of 30 preset voices, Flash TTS lets you specify a role, intonation, and timbre in natural language to create new voices across more than 100 languages and dialects. The library of ready-to-use voices has grown to more than 2,000, including regional variants such as Mexican Spanish, Quebec French, and Scottish English.

Flash TTS can also clone a specific voice from as little as a 30-second audio sample. Google has added a consent verification step, however: cloning only proceeds when a recording of the voice owner’s spoken consent matches the reference speaker. “Voice remixing,” which fine-tunes timbre, pitch, pace, and regional accent through prompts, is coming later.

Script-level direction has been added as well. You can write performance notes line by line to control pacing and emotion, and insert nonverbal expressions such as <laughs> and <sigh> or backchannel cues like |mhm|. The models can also generate a two-speaker conversation from a single script with natural turn-taking, a feature clearly aimed at podcast and audio drama production.

Benchmarks and safeguards

Google reports that Flash TTS placed first on Hume AI’s voice design benchmark with a score of 71.4, and led on regional accent rendering with 60.8. On overall quality metrics, the company says Flash TTS and Flash-Lite TTS took first and second place respectively. In blind preference testing on Voice Arena, Google says the models ranked near the top in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi. Korean results were not mentioned in the announcement.

On the safety side, all generated audio carries SynthID, an inaudible watermark that allows AI-generated audio to be detected, and voice cloning additionally applies C2PA content credentials. Voice cloning via AI Studio, however, is not available in Illinois and Texas in the United States, or in the European Economic Area (EEA), the United Kingdom, Switzerland, and India. Korea is not on that exclusion list.

For enterprise customers, API access will arrive later through Gemini Enterprise, while general users will encounter the models in Gemini Notebook (Flash TTS) and Google Vids (Flash-Lite TTS). Developer platforms including Agora, LiveKit, Pipecat, and Vercel support integration through the Gemini API, and Figma, HeyGen, and Wondercraft are already among the early adopters.

For developers in Korea, the open questions are pricing and Korean-language quality. Google did not disclose per-token rates or usage tiers in this announcement, nor did it detail the size of the Korean voice library or the extent of Korean dialect support. And because voice cloning touches on personal data and likeness rights, any company planning to build it into a commercial service in Korea will need to examine how the consent verification process lines up with local law.