Google has introduced two new text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, expanding the company’s Gemini Audio lineup with tools for custom voice creation, directed dialogue and higher-volume speech generation.
The distinction is practical. Flash TTS is positioned for creative control: users can prompt for an original voice, specify attributes such as role and accent, and direct individual lines with cues for pacing, delivery and dialect changes. Flash-Lite TTS is aimed at volume-sensitive workloads including dubbing, audio-content production and voice agents.
Both models are available now through the Gemini API and Google AI Studio. Google says enterprise API availability through Gemini Enterprise is coming soon. Flash TTS is also rolling out in Gemini Notebook, while Flash-Lite TTS is being integrated into Google Vids.
What changed: voice becomes a configurable product layer
The notable shift is not simply another library of synthetic voices. Google is offering a workflow for designing, saving and reusing a voice identity, rather than treating speech as a one-off conversion from text to audio.
According to Google, Flash TTS supports natural-language voice design in more than 100 languages and dialects, alongside a library of more than 2,000 production-ready voices. The company also says users can create a consistent vocal profile from a 30-second sample—but only for their own voice or one they have rights to use.
For builders, the model controls extend to long-form audio and two-speaker scripts. Google says the models can maintain voice quality and timbre over hours of output, stage conversations from one script, and interpret specified non-verbal cues such as laughs, sighs and acknowledgements. Those capabilities matter for teams producing localized training, narrated knowledge content, customer-service flows, podcasts or interactive product experiences.
Why it matters to operators
Voice applications often break down operationally in three places: localization, consistency and editing. A generic voice catalogue may be adequate for a prototype, but it is harder to preserve a recognizable brand voice across languages, revise individual lines without re-recording a session, or make conversational agents sound appropriately responsive without building elaborate audio pipelines.
Google’s split product strategy addresses two distinct buying decisions. Teams creating brand characters, premium narration or scripted media may value the additional direction offered in Flash TTS. Teams measuring unit economics for high-volume audio—such as support automation or localization—will be more interested in Flash-Lite’s stated focus on cost-efficient scale. Google has not detailed pricing in this announcement, so real deployment choices will still depend on API cost, latency and output reliability in production.
The rollout also places TTS inside tools where work already happens: AI Studio for experimentation, the Gemini API for applications, Notebook for user-facing audio generation and Vids for video creation. Google says platforms including Agora, LiveKit, Pipecat and Vercel support its Gemini TTS models, potentially reducing implementation work for teams already using those stacks.
Trust features are part of the product requirement
Voice replication is commercially useful and unusually sensitive. Google says replication requires a verbal consent recording from the voice owner that must match the reference speaker. It also says every clip generated by Gemini Audio models carries SynthID watermarking, and it cites C2PA credentials as part of its transparency approach.
These controls do not eliminate the need for company policy. Enterprises should still define who can approve a cloned voice, how recordings and consent evidence are retained, which markets and customer interactions are permitted, and how generated speech is disclosed. Procurement and legal teams should also test whether watermark detection and provenance information fit their own moderation, incident-response and partner-distribution processes.
What to watch next
Google cites top placements for its models on Hume AI’s Voice Design Benchmark and Overall Quality Index, as well as strong Voice Arena results in several languages. Those are useful signals, but teams should validate performance on their own scripts, accents, noisy prompts and latency requirements.
The next test is whether the models make multilingual voice products easier to operate at scale—not merely more expressive in demos. Watch for pricing details, Gemini Enterprise availability, the promised voice-remixing feature, and evidence from deployments in dubbing and conversational agents.
