MonsterHouse: An Interactive Projection on NVIDIA Nemotron Models

On October 30, 2026, a blog post by Rick Grandy and Maryam Motamedi about the MonsterHouse project was published on the NVIDIA blog. It is an interactive projection for Halloween: it listens to children who come for candy (trick-or-treaters) and responds in the voice of a grumpy old pirate ship. According to the authors' description, the character softens if visitors behave politely.

The stack is built on NVIDIA Nemotron Voice Agent. It includes Nemotron ASR (speech recognition), Nemotron 3.5 Bolt E4B at 4 GB, Magpie TTS Zeroshot (speech synthesis without separate tuning for a voice), and Audio2Face-3D (facial animation driven by sound). Everything runs on a single workstation. The difference from a typical voice demo is not only the dialogue: the responses are tied to an animated face and to a physical projection, that is, to the scene in front of the door. The article lists sections about the design of Monster House, about the "splintered grin," about teaching bad manners, about the voice, and about the monster's control panel, but there are no specific implementation details in the provided fragment.

In practice, this is an example of assembling a conversational character from ready-made open models on a single computer, without a separate cluster. Such a set can be replicated for an installation, a museum exhibit, or another interactive object. The authors also promise a section on how to build your own "unintelligent" character.

What remains unknown: the exact response latencies, the hardware requirements for the workstation, and how exactly the character determines politeness. There is no data about this in the provided fragment.