Gradium Creates Custom Synthetic Voices in Seconds

Voice agent teams have a specific problem, and it shows up the same way every time: the catalog has 400 voices, and the brief needs the one voice that isn't there. A Quebecoise receptionist for a car dealership in Montreal. A narrator in his sixties who sounds like he's held a lecture hall's attention for decades. Cloning has been the workaround, but it solves the problem one speaker at a time, and each one drags along sourcing headaches, consent paperwork and a licence to manage.

Gradium, the Paris-based voice AI company that spun out of the Kyutai research lab, is betting there's a faster route. Its new tool, Voice Design, takes a written description and hands back a finished synthetic voice in seconds, no recorded speaker, no reference audio, no rights to clear. It's built to run inside the same infrastructure teams already use: live now in the Gradium API and in Studio, free across every plan including the free tier, with any voice a team keeps running on the identical streaming Text-to-Speech endpoint as the rest of the catalog, matching latency and output formats.

The mechanics of how that description actually turns into a voice come down to what the model is trained to listen for.