Gemini Learns to Design Voices From Scratch by Description

Google DeepMind released two new speech synthesis models - Gemini 3.8 Flash TTS and a lighter Flash-Lite TTS. Now a voice is built from scratch by a verbal description of accent and character. To copy someone else's voice exactly, a 30-second recording is enough (but only with confirmed consent from the author). On the Hume AI voice design benchmark the bigger model took first place with a score of 71.4. Both versions are already available through the Gemini API and Google AI Studio.

A voice from text description removes the need for a studio and an actor if a product needs voiceover. I connect voice models into my own projects through the API, and one limitation is interesting to me. Voice cloning is not available in several regions of the world and needs explicit consent from the author (I think it's more protection against forgery than a loophole). A first place in one benchmark still doesn't mean live sound in a long dialogue - worth checking by ear before putting the model into a product.