Skip to content
context Local AI

An NVIDIA employee built the MonsterHouse Halloween installation with the Nemotron voice agent

only one source so far

An employee of NVIDIA created the MonsterHouse Halloween installation – a talking garage door playing the role of a grumpy pirate ship. It uses the Nemotron Voice Agent pipeline with the Nemotron ASR, Bolt E4B, Magpie TTS, and Audio2Face-3D models, running offline on a single workstation with RTX PRO cards.

An employee of NVIDIA created a Halloween interactive installation called MonsterHouse – a garage door with an animated mouth that reacts to visitors with the voice of a grumpy pirate ship. According to the author, the character backs down from its grumpiness if a visitor says "please," and bids a warm farewell after "thank you."

The installation runs on the Nemotron Voice Agent pipeline from NVIDIA, which combines several models: the Nemotron ASR model converts visitors' speech to text, responses are generated by the open small language model Nemotron 3.5 Bolt E4B (in 4-bit quantization, the so-called "4GB" version), the voice is synthesized by the Magpie TTS Zeroshot model, and speaking is animated by the Audio2Face-3D tool in Unreal Engine 5. For costume recognition, the author used the Nemotron 3 Nano Omni model, and for tracking and counting visitors, the YOLO and ByteTrack tools. The entire system runs offline on a single workstation with RTX PRO 6000 Blackwell Max-Q and RTX PRO 4000 Blackwell cards.

The author created the character's personality using a LoRA adapter trained on more than 5000 dialogue conversations. According to the author, more than ten models were tested, and the Bolt E4B model was chosen for its lower memory requirements compared to the Qwen3 30B-A3B and Qwen 3.6 35B-A3B models, to which the author had access even before their public release. Details can be found in the source article.

What changed

Why it matters

For developers and enthusiasts, this is a concrete, documented example of how to combine speech recognition, a small language model, voice synthesis, and 3D facial animation into a single conversational agent running offline on one workstation, without the cloud. It also shows how a LoRA adapter trained on a custom dialogue dataset can give a model a consistent personality.

Relevant practical impact

What this means

01

For individuals

Anyone looking to build their own voice AI projects gets a documented example of an offline pipeline (speech recognition, a small language model, voice synthesis, animation) running on a single GPU workstation.

What to do Review the described pipeline architecture as a reference when planning your own local voice agent.
More practical updates →
ASR Audio2Face Nemotron NVIDIA TTS Voice Agent

Check the original

Event sources

only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.

1
Hugging Face Blog primary source · first detected nemotron-halloween