Home Assistant Voice Preview + Ollama as a local AI tutor for kids

[sukumar81 (Sukumar Ravi)]

Building a Local AI Teacher with Home Assistant Voice Preview + Ollama

I recently started experimenting with Home Assistant’s Voice Preview Edition to build something a little different from the usual smart-home use cases: a local AI-powered teacher for my son.

The goal was simple: instead of an assistant that just answers questions or controls lights, I wanted something that could explain concepts, ask follow-up questions, and encourage critical thinking.

Hardware & software setup

Home Assistant side

• Home Assistant Green
• Home Assistant Voice Preview Edition
• Home Assistant Cloud (speech-to-text and text-to-speech)

AI side

• GMKtec K8 mini PC

  • Ryzen 7 8845HS
  • 32 GB RAM

• Ollama running locally

• Qwen 3 8B model

• Custom “Cambridge Grade 8 Teacher” system prompt

Architecture

Home Assistant Voice Preview
            ↓
     Home Assistant Green
            ↓
     Home Assistant Cloud
   (speech-to-text pipeline)
            ↓
     Ollama on GMKtec K8
       (Qwen 3 8B)
            ↓
     Home Assistant Cloud
   (text-to-speech pipeline)
            ↓
Home Assistant Voice Preview

Teacher persona

The model is configured as a Cambridge Grade 8 teacher that:

• Explains concepts instead of simply giving answers.
• Breaks difficult topics into smaller steps.
• Uses age-appropriate language.
• Encourages curiosity and reasoning.
• Asks follow-up questions to verify understanding.
• Avoids directly solving homework problems.

Why I built this

I wanted to explore whether Home Assistant Voice could become more than a home automation interface and evolve into something useful for everyday family life.

Everything runs locally except speech recognition and voice synthesis, so there are no external AI APIs involved.

Current limitations/challenges

• Response latency is still higher than expected (especially with reasoning-capable models).
• Conversations are not yet as fluid as ChatGPT Voice or Alexa.
• Maintaining natural back-and-forth discussions with children requires additional tuning.
• Prompt engineering has a significant impact on both response quality and speed.

The attached video was recorded immediately after the initial setup and testing, so the hardware placement and overall experience are still very much a work in progress.

I’d love to hear how others in the Home Assistant community are using Voice Preview with local LLMs, especially if you’ve found ways to reduce latency or improve conversational flow.

"Home Assistant Voice Preview + Ollama: Local AI Teacher for My Son"

#HomeAssistant
#HomeAssistantVoice
#VoicePreviewEdition
#Ollama
#LocalAI
llm
#Homelab
#Education

[NathanCu (Nathan Curtis)]

First your ask:

Friday's Party: Creating a Private, Agentic AI using Voice Assistant tools

Well, Ok I've been cajoled into it long enough. Here... I'm going to document some random thoughts I've collected about building an Agentic AI - 'Friday' in Home Assistant. (edit: and feel free to plagarize - this is how your AI disclaimer should look) Disclaimer: Friday is a Construct who is now inside MY HA install using the ZenOS-AI platform which I detail (ok rabbit hole and wander off too) below. Zenos-AI is my project which is co-authored by OpenAI Codex and Claude Sonnett 4.6. All …

Second.
I’m extremely interested in your actual prompt and constructs…

Id probably set up what you’re trying to do as a second auxiliary agent called from the main Frontline agent. There’s enough to blow the chat context for anything else and the instructions you give to not answer will directly conflict with and poison the main Frontline prompt. I’m also interested in how your isolating the curriculum corpus.

My day job is the National AI EdTech technical sales lead for a large national Ed tech provider in the US. (read: this is my entire wheelhouse) Google and MS are investing HEAVILY on this space currently and there’s a lot of lessons they have from the experiences…