Voice and local AI sizing calculator
Home Assistant Voice AI Hardware Calculator
This calculator sizes the complete Home Assistant voice stack rather than one component. It combines voice satellites, simultaneous users, speech-to-text mode, Whisper model, Ollama model class, context, latency target and power preference.
Quick answer
What this voice workload needs
Focused Speech-to-Phrase can stay on very modest hardware. Open-ended Whisper moves the minimum toward N100-class and above; adding Ollama can make system RAM or GPU VRAM the dominant requirement.
Live Amazon hardware
Current products that fit this decision
Compare current Amazon listings relevant to this guide. Product availability and prices can change.
Buying decision
Match compute to speech and LLM load
Use the result as a hardware tier, then verify language support, exact model memory, current Ollama GPU support and live product configuration. No calculator can guarantee latency across every voice model and home network.
Interactive decision tool
Home Assistant Voice AI Hardware Calculator
Size the voice and local-AI workload before comparing products. Results are planning tiers, not benchmark guarantees. Verify current Home Assistant voice support and Ollama hardware support for the exact system.
Compatibility checklist
Four checks before you purchase
Choose focused or open-ended STT first
Speech-to-Phrase and Whisper have very different compute needs.
Treat LLM memory separately
Model size, quantization, context and concurrency drive RAM/VRAM.
Keep Piper in perspective
Text-to-speech is usually lighter than Whisper or Ollama.
Measure end-to-end latency
Microphone, network, STT, LLM and TTS all affect response time.
The calculator separates speech load from LLM memory
A complete voice request can be slow because transcription is heavy, because the LLM is too large for memory, or because both happen at once. The calculator therefore scores speech mode, room concurrency and model/context separately. It does not reduce everything to one CPU benchmark. This makes the recommendation easier to interpret when a user changes from Speech-to-Phrase to Whisper or adds Ollama later.
Room count is weighted by simultaneous use
Ten satellites do not generate ten concurrent requests in most homes. The tool asks for both installed rooms and simultaneous users so a large but quiet deployment is not automatically oversized. Concurrency matters because Whisper jobs and Ollama contexts can overlap. If a household normally speaks one request at a time, a modest server may support many endpoints without difficulty.
Speech-to-Phrase keeps the compute score deliberately low
Focused local speech is designed for known Home Assistant commands and can run quickly on low-powered hardware. The calculator reflects that by keeping Speech-to-Phrase in the appliance class unless other workloads raise the requirement. This prevents a common recommendation error: selling an AI mini PC to someone who only needs private local control of lights, scenes, covers and timers.
Whisper Base starts at the N100 planning tier
Home Assistant currently recommends at least N100-equivalent hardware for Whisper Base in a fully local Voice Preview Edition setup. The calculator uses that as the open-ended speech floor and increases the tier for larger models, demanding languages, fast latency targets and simultaneous requests. It does not claim every N100 will produce identical timing, because software, thermals and audio quality still influence results.
The LLM selector estimates capacity rather than exact bytes
A 3B, 8B or 14B choice is converted into a planning tier for system RAM and GPU VRAM. Quantization varies too much for an exact universal memory number. The output is intentionally conservative and tells users to check the actual model artifact before purchase. This is more useful than a false-precision calculator that promises a model will fit based only on parameter count.
Context raises the memory tier
Home Assistant’s Ollama integration defaults to an 8K context window, and Ollama states that larger context increases memory consumption. Selecting 32K therefore adds capacity headroom in the calculator. A bigger context may help large homes or longer conversations, but users should not pay for it if 8K already works. The control exists to expose the tradeoff between context capability and hardware cost.
Latency preference changes how aggressively the tool recommends GPUs
A relaxed household can accept CPU generation for a small model, while a conversational voice assistant benefits from faster first-token and generation times. Selecting the fastest response target pushes the recommendation toward stronger CPU or supported GPU hardware. The tool never treats a GPU as mandatory simply because Ollama is enabled; it asks whether the desired user experience justifies the extra power and cost.
Power priority can favor a split architecture
When low idle power is selected alongside a larger LLM, the most efficient answer may be two machines: a tiny always-on Home Assistant/voice host and a GPU server that runs only when needed. This is why the output includes an architecture recommendation rather than merely a processor. Hardware sizing should account for hours spent idle as well as seconds spent answering a question.
Voice endpoints are displayed separately from compute
The live product section includes room endpoints because microphone and speaker quality affect the experience, but the calculator does not pretend they add server inference performance. A Voice Preview Edition or other compatible satellite captures and plays audio; the central services do the speech and model work. Keeping those roles separate helps users spend on the component that addresses the actual problem.
Storage is a capacity recommendation, not an inference score
The tool can surface 1TB–4TB NVMe options because model libraries grow with experimentation. It does not add SSD benchmark speed to the AI score. Once a model is loaded, CPU/GPU memory and compute dominate response time. Buy enough reliable local storage for models, backups and logs, then use RAM/VRAM telemetry to judge inference bottlenecks.
The result is deliberately a hardware class
The output says N100/N150, N305/Core/Ryzen, higher-end mini PC, or GPU-backed server rather than naming one permanent winner. Amazon prices and model availability change, so product cards are kept live and separate from the calculator logic. This allows Cloudzat to update the catalogue without changing the workload methodology every time a new mini PC revision appears.
Recalculate after the system evolves
Voice deployments tend to grow: one satellite becomes four, Base becomes Small for a difficult language, a 3B model becomes 8B, or context increases. The same calculator can show whether the existing host still has headroom or whether Ollama should move to another machine. Re-running the inputs is preferable to assuming the original server must remain the permanent home for every new feature.
Questions people ask
Home Assistant Voice AI Hardware Calculator questions
Can Home Assistant voice run fully locally?
Yes. Home Assistant documents fully local pipelines using local speech-to-text and Piper text-to-speech, with no requirement to send spoken commands to a cloud service.
What is the lightest local speech-to-text option?
Speech-to-Phrase is designed for focused home-control commands and can run very quickly on low-powered Home Assistant hardware.
What CPU does Home Assistant recommend for Whisper?
For fully local Voice Preview Edition use with Whisper Base, Home Assistant currently recommends at least an Intel N100 or equivalent processor.
Does Piper need a GPU?
Usually no. Piper is optimized for modest hardware and is much lighter than open-ended Whisper or a local LLM.
Does Home Assistant run Ollama itself?
The Home Assistant Ollama integration connects to an external Ollama server, which may be on the same machine or another computer reachable over the network.
How much context does Home Assistant use with Ollama?
Home Assistant currently defaults its Ollama integration to an 8K context window. Increasing context can improve capacity for larger homes but raises memory use.
How many entities should I expose to a local LLM?
Home Assistant currently recommends exposing fewer than 25 entities when experimenting with local LLM control, especially because smaller models can make more mistakes.
Is NVIDIA required for Ollama?
No. Ollama documents NVIDIA, supported AMD/ROCm, Apple Metal and experimental Vulkan paths. Compatibility depends on the exact hardware and operating system.
Should I put Ollama on a separate server?
Separate it when GPU power, memory pressure, experimentation or noise would compromise the always-on Home Assistant host. Small models can stay consolidated when simplicity matters more.
What should I measure after installation?
Measure end-to-end voice latency, Whisper transcription time, LLM first-token time, RAM/VRAM use, network delay and behavior when more than one request arrives.
Official references and methodology
Verify current Home Assistant and Ollama support
Compare current Amazon listings relevant to this guide. Product availability and prices can change.
- Home Assistant - Fully local voice assistant
- Home Assistant - Voice Preview Edition
- Home Assistant - Whisper
- Home Assistant - Piper
- Home Assistant - Wyoming Protocol
- Home Assistant - Ollama integration
- Ollama - Hardware support
- Ollama - Context length
- Ollama - Docker
As an Amazon Associate, Cloudzat may earn from qualifying purchases. Prices, seller terms, exact configurations and Home Assistant or Ollama hardware support can change.