NVIDIA PersonaPlex: Open-source full-duplex voice hits 170ms

PersonaPlex shifts conversational AI from slow, sequential turns to fluid, human-like dialogue by listening and speaking simultaneously.

• –Full-duplex architecture handles interruptions and "backchanneling" naturally, a major hurdle for current voice assistants.
• –Sub-200ms latency outperforms commercial leaders like Gemini Live, effectively removing the "thinking" pause between exchanges.
• –Open-weights distribution on GitHub and Hugging Face allows for developer fine-tuning and local experimentation.
• –High hardware floor (24GB VRAM) remains the bottleneck for local users, though 4-bit quantization may soon lower the entry barrier.