Google has quietly redefined what a text-to-speech model is supposed to be. On September 23 the company shipped Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, describing the release as a shift "from static presets into a dynamic creative studio." Both models are live now in the Gemini API and Google AI Studio, with enterprise APIs following soon; on the consumer side they land in Gemini Notebook and Google Vids.

Two models, two jobs

Flash TTS is built for deep creative direction: game characters, immersive audiobooks, podcasts, interactive media. Its headline capability is generative voice design — describing a role, accent and voice characteristics in natural language, and getting back a voice that never existed. Google's own examples include a high-energy DJ from Melbourne, a super-tinny monotone robot, and a Japanese dragon, covering more than 100 languages and dialects.

Flash-Lite TTS is the volume tier: optimized for high-volume dubbing, audio content production and expressive voice agents, while keeping fine-grained control over tone, pacing and expressive nuance.

The preset library has grown from 30 original voices to 2,000+ production-ready voices, including regional varieties like Mexican Spanish, Quebec French and Scots English. Voice replication needs just a 30-second sample — but Google hard-wired the consent step: users must provide a verbal consent recording from the voice owner, and the system verifies it matches the reference speaker before any voice is created. Replicated voices carry C2PA content credentials, and every clip from Google's Gemini Audio models is watermarked with SynthID.

Direct the performance, line by line

Both models support line-by-line direction: write stage directions for each line, or let Gemini steer delivery from natural script cues. Scripts can include non-verbal bursts like laughs, sighs and gasps rendered through inline tags, plus active-listening interjections such as "mhm" or "yeah" for comedic timing and reaction beats. A single script can stage a native two-speaker scene, with both voices kept distinctly separated and natural turn-taking. Long-form generation holds voice quality and character timbre across hours of continuous audio with minimal speaker drift, according to Google. TTS is becoming a performance instrument rather than a reading tool.

Benchmark numbers: official claims, read the attribution

The official announcement says Gemini 3.8 Flash TTS takes the #1 overall spot on Hume AI's Voice Design Benchmark with a score of 71.4, and leads accent modeling at 60.8. On Hume AI's Overall Quality Index, Flash and Flash-Lite sit at #1 and #2 respectively. Google's comparison baseline is its own Gemini 3.1 Flash TTS, with major improvements claimed in long-form content and dual-speaker screenplay control. In blind human preference runs on Voice Arena, the two models secure top positions in key languages including Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. Caveat: these are Google's own published numbers; independent third-party reproduction is still pending.

Ecosystem on day one, and a compliance map

The launch ships with an ecosystem: Agora, LiveKit, Pipecat and Vercel already integrate the models, and Figma, HeyGen, Linguana, Wondercraft, 99.co and Ollang are building on them for global dubbing, regional-accent localization and conversational voice agents at scale. On the safety side, beyond SynthID and consent verification, voice replication through AI Studio is unavailable in Illinois, Texas, the EEA, the UK, Switzerland and India — compliance boundaries drawn directly into product functionality.

So what

The TTS track has been busy: Qwen3-TTS, Bilibili's IndexTTS-2.5 and BreezeBlue's Breeze TTS 2 have all shipped from the open camp in recent months, competing on multilingual coverage, latency and emotion control. Google's move competes on none of those axes. It competes on controllability — voices designed from scratch, performances directed line by line, cloning gated by consent, outputs traceable through watermarks. For teams building audiobooks, dubbing or voice agents, the selection question is shifting from "which model reads most naturally" to "which model takes direction." The open-source camp may not lose on leaderboards, but in director-grade workflows plus compliance credentials, Google just raised the bar. What to watch next: independent reproductions on Hume and Voice Arena, and how the open community answers on controllable performance.

References:

  • Official announcement: blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/
  • Independent coverage: thenextweb.com/news/gemini-tts-3-8-flash-voice-design-cloning