NVIDIA-AI-Blueprints/nemotron-voice-agent — reverse-engineered prompt

Build me a working browser voice assistant using NVIDIA Nemotron models. I want to open a simple web page, allow microphone access, talk naturally, and hear the assistant answer back with low delay. It should stream my speech to text, send it to a Nemotron chat model, then speak the response back, and it should handle interruptions so I can cut it off and keep talking.

Please set up the backend, frontend, Docker deployment, and environment configuration so I can run it with my NVIDIA API key and NGC login. Make the setup friendly, with clear steps for local or server deployment, including the Chrome microphone permission note and any TURN server note for remote access. Include sensible defaults, comments in the env example, and a short troubleshooting section for common microphone, GPU, container, and model startup issues.

Use the current NVIDIA NIM and Nemotron docs online if you need to check details.