Intel / article

nemotron-halloween

Source: huggingface.co
Date: Published Oct 30, 2026

Why it matters

Demonstrates a full voice-agent pipeline ( ASR, small LLM, zero-shot TTS, facial animation) on a single workstation with open NVIDIA models. Useful reference for local voice agent latency and component choices.

Key takeaways · AI-distilled

The author reports a median of 1.70 seconds from end of speech to audio with Bolt Q4 and TTS, versus 0.72 seconds for cached responses, and notes that end-of-speech detection currently waits for 300 ms of silence.
The character comes from a adapter on Bolt E4B trained on over 5,000 multi-turn conversations that a local and a very large created and judged; each candidate LoRA took about one to a couple of hours to train, and the agent benchmarked them.
The author advises keeping conversation state, such as whether a visitor has already said please, in application code rather than asking the model to remember it, and training a LoRA only when prompting is not enough.
The build splits work across two GPUs, with the language and voice pipeline on an RTX PRO 6000 Blackwell Max-Q and Unreal Engine plus Audio2Face-3D on an RTX PRO 4000 Blackwell; the author says compute, not memory, is the constraint, to keep animation frame rate steady.
Because Audio2Face-3D's default weights work best on clean, neutral speech, the author feeds it the unprocessed voice and plays deep demonic effects through a separate audio stream, with about 100 ms of delay to compensate for Audio2Face-3D buffering.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.

Read the source huggingface.co