Google has launched two Gemini 3.8 text-to-speech models that let developers design voices from scratch, direct performances line by line and replicate a voice from a 30-second sample with consent checks. The important shift is from choosing a preset voice to directing an entire performance, which could change how podcasts, games, audiobooks and voice agents are made.
Google DeepMind Watch analysis
What happened
Google says Gemini 3.8 Flash TTS is built for expressive voice design, while Gemini 3.8 Flash-Lite TTS is aimed at high-volume dubbing, audio creation and voice agents. Both are available across Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook and Google Vids, according to Google’s announcement.
The Flash model supports custom voices across more than 100 languages and dialects, with controls for accent, pacing, emotion, acting cues and conversational sounds such as laughs, sighs and backchanneling. Google also says its library contains more than 2,000 production-ready voices. Voice remixing is listed as coming soon.
Key findings
- Design voices with prompts
Developers can specify a character’s role, accent and vocal qualities rather than settle for a fixed preset. - Direct every line
Scripts can include instructions for pacing, emotion, delivery and realistic conversational reactions. - Replicate with safeguards
Google says a 30-second sample can recreate a voice when the user has permission, with consent verification, SynthID watermarking and C2PA credentials. - Build longer scenes
The models support long-form audio and two-speaker dialogue while aiming to limit voice drift.
Why it matters
This makes synthetic speech feel less like typing text into a narrator box and more like directing a small cast. A game studio could create a recurring character, a publisher could produce multilingual audio, and a voice agent could respond with more control over tone instead of sounding permanently trapped in customer-service mode.
The catch is that a convincing voice is also an identity technology. Google’s consent and provenance measures matter because the same tools that help make an audiobook can make impersonation easier. The announcement describes those protections, but does not establish how reliably they prevent misuse in the wild.
Our read
This is a meaningful product step because it targets the fiddly part of speech generation: not merely pronunciation, but performance. Google’s strongest pitch is creative control at scale. Its biggest test will be whether consent, watermarking and credentials remain useful once generated audio leaves Google’s own ecosystem.
For anyone building a voice product, the practical takeaway is to test the actual workflow, not just the demo clips: check pronunciation, emotional consistency, long-form drift and the permissions trail for every replicated voice. The robot has been given stage directions. Now someone needs to check the theatre’s fire exits.
What to watch
- Whether custom voice creation and replication are broadly available or limited by region, account type or product.
- How independent users rate the models against existing voice-generation services.
- Whether watermarking and C2PA data survive editing, compression and reposting.
- When voice remixing arrives, and what additional permissions it requires.
Discussion spark: Should voice replication be allowed wherever a person has given consent, or should platforms impose stricter limits on how a replicated voice can be used?
Sources and evidence
not affiliated with, endorsed by, or operated by Google or Google DeepMind