From the source
An NVIDIA employee built MonsterHouse, a Halloween garage-door installation that talks to trick-or-treaters, using the NVIDIA Nemotron Voice Agent with Nemotron Speech, Nemotron 3.5 Bolt E4B 4GB, Magpie TTS Zeroshot and Audio2Face-3D on one workstation.
- The house listens through Nemotron Speech ASR, writes replies with Bolt E4B, voices them with Magpie TTS Zeroshot from a short reference sample, and animates a wooden mouth in Unreal Engine 5 through Audio2Face-3D.
- A LoRA adapter trained on more than 5000 multi-turn conversations supplies the grumpy pirate-ship personality, and Nemotron 3 Nano Omni describes costumes while YOLO and ByteTrack count visitors.
- The author says Bolt E4B was tested before its public release and chosen over Qwen3 30B-A3B and Qwen 3.6 35B-A3B for its smaller footprint, with the 4-bit quantized release as the default response model and the BF16 checkpoint kept for comparison.
- The build runs on a single workstation with an NVIDIA RTX PRO 6000 Blackwell Max-Q and an NVIDIA RTX PRO 4000 Blackwell, and the character follows a conversation flow where a "Please" breaks through immediately and "thank you" earns a warm farewell.
