NVIDIA Nemotron Powers a Grumpy Halloween House on a Single GPU Workstation

Engineer Rick Grandy turned his garage into a talking, grumpy pirate house for trick‑or‑treaters. The system listens with Nemotron Speech ASR, reasons with the 4 GB Nemotron 3.5 Bolt E4B model, speaks via Magpie TTS Zero‑shot using a short voice reference, and animates a wooden mouth through Audio2Face‑3D inside Unreal Engine 5. A tiny Nemotron 3 Nano Omni model even describes costumes while YOLO and ByteTrack count visitors.

Why it matters for GPU / AI infrastructure

The entire pipeline — speech recognition, language generation, text‑to‑speech, and facial animation — runs on a single workstation powered by an NVIDIA RTX PRO 6000 Blackwell Max‑Q and an RTX PRO 4000 Blackwell. This demonstrates how compact, open Nemotron models can deliver real‑time multimodal AI on modest GPU resources, lowering the barrier for interactive installations and edge deployments.

For teams building voice‑driven agents or digital avatars, the Nemotron Voice Agent stack offers a ready‑made orchestration layer that eliminates the need to stitch together disparate services. The ability to swap in a LoRA adapter for personality and to drive high‑fidelity lip sync from a single GPU makes rapid prototyping practical for marketing, retail, or entertainment projects.