Gemini Discovers How to Invent New Voices Based on Descriptions

Google DeepMind released two speech synthesis models: Gemini 3.8 Flash TTS and its compact version Flash-Lite TTS. These models are capable of generating synthetic speech from a textual specification of the accent and personality. Yes, it means that you can clone someone’s voice (well, if they agree). And you will need only a 30 second audio sample to do so! Besides, the Gemini model achieved the best result (71.4) on the Hume AI benchmark for synthetic voice design. Both models have been integrated into the Gemini API and Google AI Studio.

Text-to-speech synthesis allows you to avoid expensive recording studios and actors. I use voice models in my projects via API. For example, there is an interesting constraint — voice cloning. Voice cloning is forbidden in some regions. And you will need permission from the owner of the source voice. Even the winner of a contest doesn’t always provide high-quality results during a long conversation. Listen to the model and verify it.