Google shipped two text-to-speech models on Tuesday, gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts, both live now in the Gemini API and AI Studio. The marquee feature is voice design by prompt: describe the voice you want in plain language, across 100-plus languages, and the model builds it. It currently sits at #1 on Hume AI’s Voice Design Benchmark with a score of 71.4. There’s also a library of over 2,000 ready-made voices.
And there’s cloning. Google’s own framing: all you need is “just a 30-second audio sample of your voice or a voice you have the rights to use.”
Read that second clause again.
The rights question is a checkbox
Thirty seconds is a voicemail greeting. It’s a podcast cold open, a YouTube intro, half of a bad Zoom call. Anyone who has ever spoken publicly has already published the input.
What neither announcement describes is any mechanism that verifies the “rights to use” part. Not a consent phrase you have to read aloud, not a review queue, not a watermark you can check after the fact. The obligation sits entirely with whoever is pasting audio into an API call, which is another way of saying it sits nowhere. Google has built the capability and written down the rule, and those are two separate systems that do not talk to each other.
I don’t think this is malice. It’s the default posture of a platform company shipping into a competitive market where everyone else offers cloning too. But the gap between policy language and enforcement is the part that will generate the lawsuits, and it’s the part Google said the least about.
Flash-Lite is where the money is
The more interesting product decision is the cheap tier. Flash-Lite TTS is pitched at high-volume dubbing and voice agents, which is not a demo use case. You do not build a discount model for people playing with funny accents in a browser tab. You build it for someone localizing a back catalog into twelve languages, or running a support line where every call burns synthesis minutes.
That’s the shape of the bet: voice as infrastructure, billed like bandwidth. The prompt-design headline gets the benchmark score and the blog coverage; Flash-Lite is the one aimed at recurring spend.
Worth noting how thin the layer above these models has become. Simon Willison published a bring-your-own-key playground for the new models the same day they launched, vibe coded with GPT-6 Astra. A working interface to a brand-new speech API, built in an afternoon by one person. The client is free. The scarce things are the voice library and the appetite to host it.
For most people the 2,000 presets are the actual product, and they carry none of the licensing exposure. The cloning feature is what makes the release newsworthy and what makes it risky, and the entire safeguard is a phrase in a sentence about audio samples.