In plain English
- What happened
- On September 23, 2026, Google DeepMind introduced two new text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. The models generate custom voices from natural language prompts, replicate an existing voice from a short sample, and let a user direct pacing, emotion and non-verbal sounds line by line.
- Why it matters to your business
- If you have ever wanted a consistent voice for your business — a phone greeting, a training video, an ad read in a second language — this replaces the old stock-preset list with something you can describe or clone. It also means anyone can attempt to replicate a voice from a recording, which is why the consent and watermarking details matter before you use it on someone else's voice, including your own staff's.
- Do this
- Try the audio playground in Google AI Studio with a script you already use, such as a voicemail greeting, and listen critically before you put it in front of customers, since enterprise access is still listed as coming soon and quality can vary by language.
Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS are Google DeepMind's newest audio generation models, built to move voice generation past a fixed list of presets. The Flash version is aimed at detailed creative direction, the Flash-Lite version at high-volume, lower-cost use such as dubbing or voice agents handling many calls.
What you can actually do with it
Gemini 3.8 Flash TTS can create a voice from scratch using a plain-language description of accent, role and character, in more than 100 languages and dialects. It also draws on a library of over 2,000 production-ready voices, including regional variants like Mexican Spanish, Quebec French and Scots English.
Separately, the models support voice replication: recreating a specific person's voice from a 30-second audio sample. Once a voice is chosen or built, you can direct delivery line by line — pacing, tone, dialect shifts, even non-verbal cues like laughs or sighs — and the models are built to hold a consistent voice across long recordings such as a full podcast episode.
Consent and watermarking, not an afterthought
Voice replication requires a verbal consent recording from the voice's owner that matches the reference sample before a cloned voice can be created.
Google DeepMind says every clip generated by these models carries a SynthID watermark, woven into the audio so AI-generated speech stays detectable, along with C2PA content credentials. That matters if you plan to use a cloned voice commercially: the consent step is a checkpoint, not just a policy statement, and it applies whether you are cloning your own voice or someone else's.
Reality check
A cloned voice that sounds convincing in a demo is not the same as one that will hold up across an hour of customer calls or a full audiobook without odd pacing or accent drift. Test it on the actual script you plan to use, not a short sample, before you rely on it with customers.
Where it's rolling out, and where it isn't yet
Both models are rolling out starting September 23, 2026 for developers through the Gemini API and Google AI Studio. Gemini 3.8 Flash TTS is also available to general users inside Gemini Notebook. Enterprise access, through Gemini Enterprise, is listed as coming soon rather than available now, so a business relying on that channel should not expect it immediately.
Google also says the models rank well on independent benchmarks, placing first overall on Hume AI's Voice Design Benchmark and taking the top two spots on Hume AI's Overall Quality Index. Those are third-party evaluations cited by Google, not an independent audit, so treat them as a starting signal rather than a guarantee for your specific use case.
Developer platforms including Agora, LiveKit, Pipecat and Vercel are building on the Gemini API, and companies such as Figma, HeyGen, Wondercraft and 99.co are integrating the new TTS models for dubbing and voice agents, according to Google DeepMind. None of that requires action from a small business; it mainly signals that more tools built on top of this voice technology are likely to show up in products you may already use.
Sources
