--- title: nemotron-halloween | daily.dev description: An NVIDIA engineer built MonsterHouse, a talking garage door for Halloween powered by the open Nemotron Voice Agent stack. It uses Nemotron Speech ASR for... canonical: https://daily.dev/posts/nemotron-halloween-gmfbmmoe4 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: nemotron-halloween | daily.dev og:description: An NVIDIA engineer built MonsterHouse, a talking garage door for Halloween powered by the open Nemotron Voice Agent stack. It uses Nemotron Speech ASR for... og:url: https://daily.dev/posts/nemotron-halloween-gmfbmmoe4 og:image: https://api.daily.dev/og/posts/gMfbmmOE4.png og:image:alt: nemotron-halloween og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # nemotron-halloween **[Hugging Face](https://daily.dev/sources/huggingface)** · 12 min read · 0 upvotes · 0 comments ## Summary An NVIDIA engineer built MonsterHouse, a talking garage door for Halloween powered by the open Nemotron Voice Agent stack. It uses Nemotron Speech ASR for listening, the 4-bit quantized Nemotron 3.5 Bolt E4B 4GB model as the conversational brain, Magpie TTS Zeroshot for voice, and Audio2Face-3D to animate a custom mouth rig built in Maya and Unreal Engine 5. A LoRA adapter, trained on a dataset of over 5,000 multi-turn conversations, gives the character its grumpy pirate personality. The system runs on two GPUs, an RTX PRO 6000 Blackwell Max-Q for language and voice, and an RTX PRO 4000 Blackwell for graphics. Median end-of-speech-to-audio latency measured 1.70 seconds with live generation versus 0.72 seconds for cached responses. An optional Nemotron 3 Nano Omni vision model identifies visitor costumes, tested on a video of five costumed sheep. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is the end-to-end latency for a voice agent using Nemotron Bolt E4B and TTS compared to using cached audio responses? Median end-of-speech-to-audio latency measured 1.70 seconds when generating a live response with the quantized Bolt E4B model and Magpie TTS, versus 0.72 seconds when using cached, pre-recorded responses. Speech-end detection also waits for 300 ms of silence before processing begins, adding to perceived response time. _Developers building responsive voice agents can track latency benchmarks like these on daily.dev._ ### How does NVIDIA Nemotron 3.5 Bolt E4B compare to Qwen3 30B-A3B for local conversational agents? Nemotron 3.5 Bolt E4B has a smaller footprint than Qwen3 30B-A3B and Qwen 3.6 35B-A3B while still being a solid choice for local conversational agents. It is an open small language model purpose-built for interactive characters, and is available as a 4-bit quantized 'E4B-4GB' release or a BF16 checkpoint. _Anyone weighing small open models for on-device agents can follow comparisons like this on daily.dev._ ### How much training data is needed to fine-tune a LoRA adapter for a conversational AI character's personality? A dataset of over 5,000 multi-turn conversations was used to train LoRA adapters for giving a conversational model a consistent personality. Some LoRA training runs took about an hour, others a couple of hours, and a local agent benchmarked results while a separate large LLM created and judged the dataset. _Engineers planning LoRA fine-tuning for custom agent personas can find practical examples like this on daily.dev._ ## Similar posts on daily.dev - [Building Voice Agents with NVIDIA Open Models](https://daily.dev/posts/building-voice-agents-with-nvidia-open-models-bhrohy86v) · Hacker News · 1 upvotes · 1 comments - [Building NVIDIA Nemotron 3 Agents for Reasoning, Multimodal RAG, Voice, and Safety](https://daily.dev/posts/building-nvidia-nemotron-3-agents-for-reasoning-multimodal-rag-voice-and-safety-kmj8nolvr) · NVIDIA Developer · 1 upvotes · 0 comments - [Voice AI in vogue: Nvidia releases new open transcription model](https://daily.dev/posts/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model-dotdulayg) · TechCrunch · 0 upvotes · 0 comments - [Scaling Real-Time Voice Agents with Cache-Aware Streaming ASR](https://daily.dev/posts/scaling-real-time-voice-agents-with-cache-aware-streaming-asr-v2urtcsgr) · Hugging Face · 15 upvotes · 0 comments --- Tags: [#unreal-engine](https://daily.dev/tags/unreal-engine), [#text-to-speech](https://daily.dev/tags/text-to-speech), [#voice-ai](https://daily.dev/tags/voice-ai) [View this post on daily.dev](https://daily.dev/posts/nemotron-halloween-gmfbmmoe4) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"nemotron-halloween","url":"https://daily.dev/posts/nemotron-halloween-gmfbmmoe4","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/nemotron-halloween-gmfbmmoe4"},"datePublished":"2026-10-30T18:02:30.301Z","dateModified":"2026-10-30T18:03:04.831Z","description":"An NVIDIA engineer built MonsterHouse, a talking garage door for Halloween powered by the open Nemotron Voice Agent stack. It uses Nemotron Speech ASR for...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/807cef05ac4aa950d8f239fd7178c3b0?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/807cef05ac4aa950d8f239fd7178c3b0?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hugging Face","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hugging Face","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/f1f55c67d81a4330acf5b90b26b0c8e1","url":"https://daily.dev/sources/huggingface"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/nemotron-halloween-gmfbmmoe4","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"unreal-engine,text-to-speech,voice-ai","timeRequired":"PT12M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hugging Face","item":"https://daily.dev/sources/huggingface"},{"@type":"ListItem","position":3,"name":"nemotron-halloween"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/nemotron-halloween-gmfbmmoe4#faq","mainEntity":[{"@type":"Question","name":"What is the end-to-end latency for a voice agent using Nemotron Bolt E4B and TTS compared to using cached audio responses?","acceptedAnswer":{"@type":"Answer","text":"Median end-of-speech-to-audio latency measured 1.70 seconds when generating a live response with the quantized Bolt E4B model and Magpie TTS, versus 0.72 seconds when using cached, pre-recorded responses. Speech-end detection also waits for 300 ms of silence before processing begins, adding to perceived response time. Developers building responsive voice agents can track latency benchmarks like these on daily.dev."}},{"@type":"Question","name":"How does NVIDIA Nemotron 3.5 Bolt E4B compare to Qwen3 30B-A3B for local conversational agents?","acceptedAnswer":{"@type":"Answer","text":"Nemotron 3.5 Bolt E4B has a smaller footprint than Qwen3 30B-A3B and Qwen 3.6 35B-A3B while still being a solid choice for local conversational agents. It is an open small language model purpose-built for interactive characters, and is available as a 4-bit quantized 'E4B-4GB' release or a BF16 checkpoint. Anyone weighing small open models for on-device agents can follow comparisons like this on daily.dev."}},{"@type":"Question","name":"How much training data is needed to fine-tune a LoRA adapter for a conversational AI character's personality?","acceptedAnswer":{"@type":"Answer","text":"A dataset of over 5,000 multi-turn conversations was used to train LoRA adapters for giving a conversational model a consistent personality. Some LoRA training runs took about an hour, others a couple of hours, and a local agent benchmarked results while a separate large LLM created and judged the dataset. Engineers planning LoRA fine-tuning for custom agent personas can find practical examples like this on daily.dev."}}]} ```