This New AI Speaks in Real Time, Making Voice Assistants Sound Human

AI voices have come a long way, but most still feel robotic or sluggish. A new system called GEPARD, from a team at Nineninesix, Inc., aims to fix that. GEPARD is a text-to-speech model that generates spoken audio the moment text arrives — no waiting for a full sentence. The result is real-time dialogue that feels like talking to a person, not a machine.

What makes GEPARD special is its speed. For a single conversation, it generates speech roughly 15 times faster than real time. When hundreds of people talk to it at once — up to 256 simultaneous conversations — it stays fast enough on one powerful GPU chip. That could let companies run natural-sounding voice agents for large numbers of customers without needing massive server farms, which could lower costs and reduce delays.

The team also tackled a quirky problem: AI speech systems sometimes drop very short words like "hi" or "no" when generating quickly. They invented a fix to keep those tiny but important words intact. GEPARD can also clone a voice from a short sample, meaning characters or assistants could speak in a chosen voice. The code and open-weight models are being released, so others can build on the work.

For everyday people, this could mean voice assistants that don't make you repeat yourself, phone bots that understand interruptions, and video games where characters respond instantly. But voice cloning also raises concerns about impersonation and scams, so it's a tool that needs careful handling.