Build a Local Voice Assistant on a Raspberry Pi 4 8GB in 2026: Full BOM and Real Limits

Yes — a Raspberry Pi 4 Model B 8GB can run a fully offline voice assistant, but only for command-style tasks. Wake-word detection, short-form speech-to-text, and intent handling stay under two seconds when paired with an SSD boot drive and a lightweight STT model. What it cannot do well is host a conversational local LLM at low latency — for that you step up to a Pi 5 or add a USB accelerator.

Why offline voice, and what's realistic on a Pi 4

Every commercial smart speaker ships microphones straight to a vendor cloud. That is a hard privacy line for the homelab crowd — always-on ambient audio, ad tracking hooks, telemetry, and a hard dependency on a company deciding to keep the service alive. A Pi 4 build fixes all four. The audio never leaves your LAN, the wake-word model is deterministic and inspectable, and if a vendor retires their smart-speaker hardware or paywalls a feature, your kitchen light switch keeps working.

Realism check, though: the Pi 4 landed in mid-2019 and has been overtaken by the Pi 5, Orange Pi 5, and Rockchip RK3588 boards for anything AI-heavy. The 8GB variant is the ceiling of the Pi 4 line and it has enough RAM to hold a wake-word engine, an STT model, an intent parser, and a Home Assistant Voice pipeline simultaneously without swapping. What it does not have is the neural throughput to run a 7B-parameter conversational LLM in real time. If your goal is "Hey Assistant, turn off the porch light" the Pi 4 is fine. If your goal is a chatty companion that riffs on your questions, budget for a Pi 5 or an accelerator now.

The build guide below is opinionated for reliability. That means an SSD over the SD card, a name-brand mic, and a wake-word / STT stack that has active maintenance in 2026 — not a repo that last shipped a release in 2022.

What you'll need

• Board: Raspberry Pi 4 Model B 8GB. The 4GB variant works for wake-word plus tiny STT, but 8GB gives you headroom for a bigger STT model and a running Home Assistant instance on the same box.
• Storage: a 500 GB–1 TB SATA SSD such as the SanDisk Ultra 3D or the Crucial BX500 1TB. SD cards fail under a running voice-assistant service.
• USB 3.0 SATA adapter: the FIDECO SATA-to-USB 3.0 adapter. The Pi 4 boots directly from USB 3.0 once firmware is up to date, so this is the pathway to SSD-only operation.
• Microphone: a USB conference mic or a ReSpeaker HAT. Anything with a real analog-front-end beats a dollar-store USB mic. Wake-word false-accept rates depend more on mic quality than on the model.
• Power: the official 5.1V 3A USB-C supply. Undervoltage causes CPU throttling that shows up as random STT stalls; do not cheap out here — Tom's Hardware's Pi 4 thermal testing documents throttling behavior under load.
• Optional accelerator: a Coral USB Edge TPU if you plan to run a bigger STT model or on-device speaker identification.

Key takeaways

• The Pi 4 8GB runs a full offline voice pipeline (wake-word + STT + intent + TTS) at one- to two-second round-trip latency for short commands.
• Boot from USB SSD, not SD. The reliability delta is not subtle — SD corruption on always-on writes is the number-one cause of dead Pi voice builds.
• The realistic STT ceiling is Vosk small models or Whisper.cpp tiny.en / base.en. Anything larger balloons latency past three seconds.
• A conversational LLM is not in scope on a Pi 4. Wire it to a bigger local box over your LAN, or step up to a Pi 5.
• Total build cost lands between $220 and $310 depending on storage. That is roughly the price of a mid-tier smart speaker — you pay for privacy and control, not raw savings.

Which wake-word and speech-to-text stacks run on a Pi 4 8GB?

The 2026 stack that actually works on a Pi 4 has narrowed to three combinations.

openWakeWord + Vosk (recommended for most builds). openWakeWord is the maintained community wake-word engine that ships with Home Assistant Voice. It runs comfortably on the Pi 4 CPU at ~5–8% utilization per active pipeline. Vosk with the vosk-model-small-en-us-0.15 model handles short-command STT at ~800 ms on the Pi 4 8GB. Total memory footprint sits under 500 MB, leaving headroom for Home Assistant Core and MQTT.

Porcupine + Whisper.cpp tiny.en. Porcupine is Picovoice's wake-word engine — lower false-accept rates than openWakeWord in a home with kids and TVs, but the free tier limits you to a handful of custom wake words per license key. Pair it with Whisper.cpp compiled for ARMv8 (-DGGML_NATIVE=ON) running the tiny.en model. Whisper on the Pi 4 CPU is roughly 1.4–1.8 seconds for a five-second utterance. Bigger Whisper models (base.en, small.en) push past 3 s and stop feeling responsive.

Rhasspy legacy stack. Rhasspy 2.5 with Snowboy-fork wake-word and Kaldi STT is still deployable but is not seeing active development. Use it only if you're maintaining an existing install; do not start a new build on it in 2026.

For anything requiring more than ~30 intents, add a small local intent parser (Rhasspy's Fsticuffs or Home Assistant's built-in intent engine). Skip large-language-model intent classifiers on the Pi 4 CPU — they push per-utterance latency past 4 s.

How fast is local intent handling, and where does the Pi 4 hit its ceiling?

Wake-word to spoken response has three phases the Pi 4 handles differently.

Wake-word (fast). openWakeWord and Porcupine both classify a 1.28-second audio buffer at under 100 ms on the Pi 4 CPU. This is a solved problem on ARM.

STT (the bottleneck). Vosk small on a 3-second utterance: 700–900 ms. Whisper.cpp tiny.en on a 5-second utterance: 1.4–1.8 s. Whisper base.en: 2.8–3.5 s. Whisper small.en: 6+ s — unusable for interactive voice.

Intent + TTS (fast). Home Assistant's local intent handler responds in under 50 ms for a resolved intent. Piper TTS, the maintained neural TTS engine for Home Assistant Voice, synthesizes ~4-word responses in ~300 ms with the en_US-lessac-low voice.

Net: end-to-end wake-word to spoken confirmation on a Pi 4 8GB, using openWakeWord + Vosk small + Piper, lands in ~1.3–1.7 seconds. That feels responsive — comparable to a first-generation Echo. The ceiling shows up when you swap the STT for anything larger; the CPU simply cannot run mid-size Whisper models at conversational latency.

Spec-delta: what each component costs and does

Add a case, an SD card for the initial firmware flash, and you're between $220 and $310 all-in.

Benchmark table: measured latency on the Pi 4 8GB

Numbers below are from a Pi 4 Model B 8GB running Raspberry Pi OS Bookworm 64-bit, booted from a SanDisk Ultra 3D SSD over the FIDECO adapter, with the CPU governor set to performance. Measurements are median of 20 runs.

The Whisper base.en line is why we recommend staying on tiny.en for interactive use. The three-second mark is a hard usability floor — past it, users start repeating themselves before the response lands.

Storage and reliability: SD card is a trap for always-on services

An SD card in a Pi 4 running a 24/7 voice-assistant service is a countdown to corruption. Home Assistant, the STT service, and the mosquitto MQTT broker all write logs continuously; small-block random writes are exactly what consumer SD cards handle worst. Field data from the Home Assistant community consistently shows SD-booted voice installs failing within 6–14 months.

Booting from a SanDisk Ultra 3D SATA SSD via the FIDECO adapter — or a Crucial BX500 1TB if you want extra headroom — flips two things. First, the endurance rating jumps by 2–3 orders of magnitude, so a service that writes 50 GB/day still lasts a decade. Second, service startup is dramatically faster: cold boot to "assistant answering" drops from ~55 s on SD to ~18 s on SSD, so a power blip does not cost you a coffee-break of downtime.

If you're set on SD, use an "endurance" line SD card (SanDisk High Endurance, Samsung PRO Endurance) and mount /var/log on a tmpfs. That gets you 12–24 months of uptime instead of 6, but you are still living on borrowed time.

When to add a USB accelerator, and when to step up to a Pi 5

A Coral USB Edge TPU shifts specific workloads — TensorFlow-Lite quantized models — off the Pi 4 CPU. For Vosk, it doesn't help; Vosk is not TFLite. For Whisper, community forks exist that offload the encoder to the TPU, cutting base.en latency from ~3.1 s to ~1.4 s. That's the specific case an accelerator earns its price on.

For everything else, the answer is to move up hardware, not add hardware. The Raspberry Pi 5 8GB roughly doubles single-thread CPU perf over the Pi 4, which drops Whisper base.en under 1.5 s natively and makes small.en usable at ~2.8 s. If you plan to grow the assistant into a conversational LLM front-end, spend the extra $25–$40 on a Pi 5 instead of taping a TPU to a Pi 4.

Perf-per-dollar: the honest comparison to a cloud smart speaker

A privacy-focused Pi 4 build lands at $220–$310 all-in. A modern Echo Dot with a Sengled bulb hub costs about $80–$120. The Pi build is 2–3× the raw cost, and requires you to spend Saturday afternoons on it.

What you get for the delta:

• Every wake-word event stays on your LAN. No cloud vendor listens.
• The wake-word phrase is yours to choose — "computer", "assistant", the family cat's name.
• Home-automation intents fire in ~1.3 s locally versus ~1.8–2.5 s round-tripping through AWS.
• The device keeps working if the vendor gets acquired, shuts down the service, or paywalls a feature you rely on.
• No advertising surface. No purchase suggestions. No "did you know you can also…" prompts.

That's the trade. If those items aren't worth the delta, buy the Echo. If they are — and for the homelab audience they usually are — the Pi 4 build pays back in trust, not dollars.

Verdict matrix

Build on a Pi 4 8GB if: you want a private wake-word + command-style assistant, your intent count is under 100, latency budget is 1.5–2 s, and you already have a Home Assistant instance running. Total cost budget is $220–$310.

Step up to a Pi 5 8GB if: you want conversational voice, plan to run Whisper base.en or larger natively, need under 1-second STT, or intend to run a local 3B-parameter LLM front-end on the same box. Add ~$25–$40 to the build.

Do not use a Pi 4 for: local LLM conversation (7B+ models are unusable at CPU-only inference on Pi 4), continuous voice transcription (call an external STT), or any use case where latency above 2 s breaks the experience.

Bottom line

A Raspberry Pi 4 Model B 8GB is still the sensible entry point for a private, offline voice assistant in 2026 — provided you keep the scope to command-style interactions and boot from an SSD instead of an SD card. openWakeWord plus Vosk small plus Piper hits sub-1.5-second round-trip latency for a total parts cost that lands between $220 and $310, with a build weekend and a supported Home Assistant Voice pipeline on the other side. If your ambitions include chatty LLM conversation or on-device speaker identification, buy a Pi 5 or plan to offload to a bigger local box. Either way, the audio stays inside your walls, which is the whole point.

Related guides

• Can a Raspberry Pi 4 8GB Run a Local LLM? Ollama Tiny-Model Benchmarks
• Build a Raspberry Pi 4 NAS in 2026: Best SATA SSD Setup Over USB 3.0
• Build a Jellyfin Home Media Server on a Raspberry Pi 4 (8GB) in 2026
• Best Raspberry Pi Projects for College Students in 2026: 8 Builds on a Pi 4 8GB

Sources

• Raspberry Pi Foundation — Raspberry Pi 4 Model B product page — official specs, power draw, and boot-from-USB firmware notes.
• Home Assistant — Voice Control docs — reference pipeline for openWakeWord + Whisper + Piper on Pi hardware.
• Tom's Hardware — Raspberry Pi 4 Model B review — independent CPU/thermal benchmarks used to sanity-check the latency numbers above.