Google DeepMind Puts Voice Design in Developers' Hands With Gemini 3.8 TTS

Google DeepMind has released two new text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, that shift voice generation away from fixed preset catalogs toward voices built on demand from natural language prompts. According to the announcement, written by Group Product Manager Leland Rechis and Alan Cowen, Director of Research Science, on behalf of the Gemini Audio team, the models are rolling out starting today through the Gemini API and Google AI Studio, with enterprise API access via Gemini Enterprise described only as coming soon. The company did not disclose pricing, latency figures, rate limits or context limits for either model.

What is actually new

The headline capability is generative voice design. Google says developers can describe a voice in plain language — a role, an accent, a set of vocal characteristics — and get a usable speaker back, across more than 100 languages and dialects. The post illustrates this with examples including a high-energy Melbourne DJ, a tinny monotone robot and a Japanese dragon, though the underlying audio samples are hosted in the original post rather than described in detail.

That is paired with a much larger stock catalog. Google frames the jump as moving from 30 original voices to an "infinite library," with more than 2,000 production-ready voices available, including regional varieties the company lists as Mexican Spanish, Quebec French and Scots English. Voice replication builds a consistent vocal profile from a 30-second reference sample. A remixing feature — adjusting timbre, pitch, pace and accent on an existing library voice, including through prompts like adding a subtle accent or softening delivery — is announced as coming soon, with no date given.

The split between the two models follows a familiar pattern. Flash TTS is positioned for creative direction and character work in games, audiobooks, podcasts and interactive media. Flash-Lite TTS is positioned for high-volume, cost-efficient workloads such as dubbing and voice agents. Google says both support fine-grained control over tone and pacing, but voice creation from scratch is described specifically as a Flash TTS capability. Teams evaluating Flash-Lite for character work should confirm what carries over, because the post does not spell that out.

Directing a performance, not just reading text

The second cluster of features targets a long-standing weakness in synthetic speech: it reads text, it does not act it. Google says both models accept line-by-line stage directions, or can infer delivery from script cues, to move between something like a calm support agent and a whispered suspense scene.

Three supporting capabilities matter more than they sound. Long-form generation is meant to hold timbre and pacing steady across hours of continuous audio with what Google calls minimal speaker drift — the failure mode that has made AI-narrated audiobooks tedious to produce, because drift forces re-generation and manual stitching. Native two-speaker scene staging produces multi-turn dialogue from a single script while keeping the voices distinct and turn-taking natural. And scripted vocal bursts allow inline non-verbal markup, using tags such as <laughs>, <sigh> and <gasp> alongside backchanneling interjections like |mhm| and |yeah|.

The tag syntax is a small detail with real workflow consequences. Inline markup means scripts become a control surface, which is good for version control and reproducibility, but it also means existing text pipelines need an escaping and validation layer before user-supplied text hits the model. Google did not describe how the models behave when tags are malformed or when untrusted input contains them.

The benchmark claims, and what they leave out

Google reports that Gemini 3.8 Flash TTS took the top overall spot on Hume AI's Voice Design Benchmark with a score of 71.4, and led accent modeling at 60.8. It also says the two models took first and second place on Hume AI's Overall Quality Index, and that in blind human preference testing on Voice Arena they secured top positions in Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. Compared with Gemini 3.1 Flash TTS, Google claims major improvements on long-form content and dual-speaker screenplay control.

These are third-party evaluations, which is better than a vendor's internal scorecard, but the post does not explain the benchmark methodology, the competitor set, the scoring scale, the size of the margins or when the runs took place. A score of 71.4 on a 100-point scale also implies substantial headroom remains. "Top positions" in the Voice Arena results is doing quiet work as well — it is not the same as first place in each language, and Google does not say which is which. For buyers, the practical takeaway is that these numbers justify a bake-off, not a procurement decision.

Consent and watermarking are built in — and geographically uneven

Google says voice replication requires a verbal consent recording from the voice owner that matches the reference speaker before a cloned voice can be created. Every clip generated by Gemini Audio models carries a SynthID watermark woven into the audio itself, and replicated voices are also backed by C2PA content credentials. The company points to a model card for further safety detail.

This is a meaningfully stronger default than an unchecked upload box, and the speaker-matching requirement raises the cost of casual misuse. The open questions are about resilience and verification. Google does not say how the watermark survives re-encoding, compression, speed changes or re-recording through a speaker and microphone, nor who can run SynthID detection — a watermark only reduces harm if the platforms hosting audio can actually check for it. Nothing in the announcement addresses how consent recordings themselves are validated against impersonation.

There is also a notable footnote: voice replication in AI Studio is unavailable in Illinois, Texas, the EEA, the UK, Switzerland and India. Google gives no reason. That list maps closely onto jurisdictions with biometric privacy statutes or strict data protection regimes, and it means any product plan assuming global cloned-voice availability needs a fallback for those markets.

Distribution is the strategic part

Beyond the API, Google says Flash TTS reaches general users through Gemini Notebook and Flash-Lite TTS through Google Vids, with Gemini Enterprise access still pending. On the infrastructure side, the company names Agora, LiveKit, Pipecat and Vercel as developer platforms through which the models can be deployed, and lists Figma, HeyGen, Linguana, Wondercraft, 99.co and Ollang as companies integrating the models for dubbing, localization and voice agents. The nature, scale and commercial terms of those integrations were not disclosed.

The models join a fast-expanding Gemini Audio line that Google says already includes 3.5 Live Translate, 3.5 Transcribe, 3.8 Live and 3.8 Live Extended Thinking. That breadth is a competitive asset, but it also raises a versioning burden for teams: audio stacks now span transcription, translation, live conversation and synthesis, each on its own release cadence.

What to watch

• Pricing and latency for both tiers, which Google did not disclose. Cost per character or per second, and time-to-first-audio, will decide whether Flash-Lite is genuinely viable for real-time agents and bulk dubbing.
• Whether the "minimal speaker drift" claim holds across multi-hour audiobook runs in practice, and what the failure mode looks like when it does not.
• A firm date for Gemini Enterprise API availability and for the voice remixing feature, both currently listed only as coming soon.
• Independent verification of SynthID robustness in audio, plus clarity on who outside Google can run detection.
• Whether the replication restrictions in Illinois, Texas, the EEA, the UK, Switzerland and India ease over time, or become the template for how cloned-voice features ship regionally.