Google has a new speech model. Voices no longer have to be picked from a list — you can describe one in words — and the safeguards against impersonation ship in the same box.
Google released Gemini 3.8 Flash TTS, a text-to-speech model, alongside a cheaper Flash-Lite TTS. With Flash TTS, developers can design a voice from scratch by describing it in plain language. It can also recreate a specific person's voice from 30 seconds of audio, but only after a consent recording from that person. It covers more than 100 languages and lets users direct performance and pacing line by line. It is available through the Gemini API and Google AI Studio, and is coming to products such as Google Vids.
Until now, synthetic speech mostly meant choosing from a set of stock voices. Being able to order a voice by description could cut the effort of producing narration or dubbing. The flip side is that cloning a voice from a short clip is exactly what impersonation scams need. Google pairs the consent check with SynthID, an inaudible watermark added to every generated clip — convenience and misuse controls built into the same product from day one.
Whether the consent check holds up without loopholes, and whether other platforms and social networks start reading the watermark and surfacing it to users. A watermark only works once the detectors are in place.