Source time Edition date
IndustryHugging Face
nemotron-halloween
Nemotron 万圣节
Nemotron's Halloween-themed release is worth a look only if you can judge whether the theming signals a real model update or just seasonal marketing.
Judgment
A Halloween display builder deciding whether to copy MonsterHouse should treat it as a reusable single-workstation voice-agent stack, not seasonal theming: the cost is one RTX PRO 6000 Blackwell Max-Q plus an RTX PRO 4000 Blackwell and a custom animation rig, so commit only if you will reuse the pipeline after October.
Basis
Rick Grandy and Maryam Motamedi, both at NVIDIA, describe MonsterHouse as an interactive projection built on the NVIDIA Nemotron Voice Agent. The pipeline is: Nemotron 3 ASR (nemotron-speech-streaming-en-0.6b) turns visitor speech into text; Nemotron 3.5 Bolt E4B 4GB, an open small language model, writes each reply, with a LoRA adapter supplying the personality; Magpie TTS Zeroshot voices the reply from a short reference sample; Audio2Face-3D animates the mouth in Unreal Engine 5. Nemotron 3 Nano Omni describes costumes, while YOLO and ByteTrack count and track visitors. It runs on a single workstation with an RTX PRO 6000 Blackwell Max-Q and an RTX PRO 4000 Blackwell. The author says the system had to be off-grid and completely self-contained because of a memory and compute budget.
- Listening ASRtext
- Bolt E4Breply
- Magpie TTSspeech
- Audio2Face-3Danimation
- Unreal Engine 5
Unclear The source does not state end-to-end latency, power draw, total hardware cost, or whether the MonsterHouse LoRA adapter and Voice Agent configuration are publicly released.
Next Next, verify whether NVIDIA's Nemotron Voice Agent materials include the MonsterHouse LoRA adapter and a measured latency figure for the two-RTX workstation.