NVIDIA PersonaPlex opens full-duplex voice AI

This is less a demo than a signal that voice AI is moving past stitched-together STT/LLM/TTS pipelines into end-to-end streaming systems. NVIDIA is betting that latency, interruption handling, and persona control matter more than generic chatbot polish.

• –Full-duplex behavior is the real differentiator: the model can listen and speak simultaneously, which is what makes conversational turns feel human instead of mechanical
• –Open weights and local deployment make it attractive for private assistants, embedded voice apps, and teams that can’t route audio through third-party APIs
• –The tradeoff is obvious: this is still GPU-heavy research-grade infrastructure, not a casual consumer app
• –The Moshi base and streaming audio architecture put it in the same broad race as other low-latency voice agents, but NVIDIA’s distribution and hardware story give it more reach than most
• –If the benchmarks hold up in practice, this could become a reference stack for real-time voice UX in gaming, support, and robotics