MonsterHouse combines voice AI with a talking garage door

Rick Grandy’s account describes a garage-door projection that listens and answers as a grumpy pirate character. Speech recognition converts visitors’ words into text; Bolt generates replies, Magpie speaks them and Audio2Face animates the mouth. Grandy works at NVIDIA. His custom MonsterHouse code is not packaged for release, and he says complete performance testing still needs real visitors.

The public Voice Agent blueprint gives developers a starting point: its Generic Assistant example connects speech recognition, a language model and speech synthesis in an English-only conversation. The software handles streaming and interruptions. Its deployment choices matter when considering a prototype. The cloud recipe calls remote model services without requiring a local GPU; workstation recipes have hardware and memory requirements that depend on the selected models. NVIDIA directs developers to each example’s own documentation for its supported recipes and configuration.

For the animated character, Audio2Face documentation describes a separate job: turning recorded or streamed audio into facial movement. In Unreal Engine, visual Blueprint nodes connect animation data to a character’s facial controls. This integration requires both the core ACE Unreal plugin and the Audio2Face Models plugin. The supplied sample demonstrates a MetaHuman setup, and inference can run locally or on a separate server.