openhome mini — DIY Local "Alexa" Build Guide

A privacy-first smart speaker on a Raspberry Pi 4 that matches an Amazon Echo's core

experience — far-field "Alexa" wake word, natural voice replies, smart-home control,

timers, music, and a light ring — but runs locally and answers only to you.

Status: design v1 (2026-07-30). Researched against current parts + software.

Companion files: hardware/enclosure/openhome-mini.scad (printable shell),

hardware/mockups/06-photoreal.jpg (retail concept), 07-mvp-prototype.jpg (build).

1. Goal & requirements (what "match a real Alexa" means)

Echo capability	openhome mini equivalent	Met by
"Alexa" wake word, hears across a room	4-mic far-field array + openWakeWord `alexa`	XVF3800 + OWW
Understands speech	Local STT (Whisper)	faster-whisper
Talks back naturally	Local TTS	Piper
Controls smart home	Home Assistant Assist	HA
Timers / alarms / weather	Assist intents	HA
Plays music	Music Assistant → speaker	HA + MA
Light-ring feedback	16-LED addressable ring	WS2812
Private / offline	nothing leaves the LAN	(Alexa can't do this)
Open-question "brain" (optional)	LLM conversation agent	Gemini/OpenAI/local

Hard requirements

Wake from ~4–5 m in a normal room, with music playing (barge-in).
End-to-end "Alexa, turn on the lights" → action + spoken confirm in < 3 s.
Single 3D-printed device, USB-C powered, serviceable, no cloud account.

2. System architecture

                       ┌─────────────────────────── openhome mini (one Pi 4) ───────────────────────────┐
  voice ──▶ [XVF3800 4-mic array] ──USB(16k mono, AEC/beamformed)──▶ ┐
                                                                     │
                                              ┌── openWakeWord ("alexa")   ── wake ──┐
                                              │                                       ▼
                       [ Home Assistant + Assist pipeline ] ── STT (faster-whisper) ─▶ intent ─▶ TTS (Piper)
                                              │                                       │
   speaker ◀─ [3" driver] ◀─ [MAX98357A I2S] ◀── audio out ◀───────────────────────┘
                                              │
   ring   ◀─ [WS2812 ×16] ◀── SPI (GPIO10) ◀──┘   (state: idle/listen/think/speak/mute)
                       └───────────────────────────────────────────────────────────────────────────────┘

Two deployment shapes:

A. Self-contained (recommended to start): everything above runs on the one Pi 4.

Simplest; ~2–4 s latency with whisper-base. This is "one box I can build."

B. Satellite + server (performance path): Pi 4 runs only mic + wake word + speaker +

LEDs (via linux-voice-assistant); a separate HA box does STT/TTS/intent. Sub-second,

scales to many rooms. Same hardware, just moves the heavy compute off-device.

3. Bill of materials (with links)

#	Part	Why	Price	Link
1	Raspberry Pi 4 Model B (4 GB)	compute / runs everything	~$55	buy
2	SanDisk Extreme 32 GB microSD (A2)	OS + models	~$10	buy
3	ReSpeaker XVF3800 USB 4-Mic Array	far-field, HW AEC/beamforming/DoA, USB or I2S	$60.99	buy
4	Adafruit MAX98357A I2S amp	digital audio out, 3 W	$5.95	buy
5	3" 4 Ω 3 W speaker	room-filling driver	$1.95	buy
6	NeoPixel Ring — 16 × WS2812	light-ring feedback	~$9	buy
7	Official USB-C PSU 5.1 V / 3 A	power (see §7 audit)	~$8	buy
8	Heatsink + 30 mm fan	Whisper heats the Pi	~$6	buy
9	M2.5 heat-set inserts + screws	serviceable base	~$8	buy
10	Dupont/JST jumper wires	wiring	~$7	buy
11	3D-printed shell (PETG, ~150 g)	enclosure (`openhome-mini.scad`)	~$3	in this repo

Total ≈ $175 @ qty 1 (~$150 at a batch of 10). Compare: Echo (4th gen) ~$100 but cloud-bound

and no local control; openhome mini adds privacy + a scriptable brain.

Budget swaps

Cheaper mic: ReSpeaker XVF3000 4-Mic ($64) or 2-Mic ($30-ish, near-field only).
No-solder audio: a small USB speaker instead of MAX98357A+driver (loses the clean I2S path).
Fancier sound: swap #5 for a Dayton Audio 2"/2.5" full-range + a small passive radiator.

4. Wiring

4a. MVP wiring (simple)

XVF3800 ── USB ─────────────▶ Pi 4 USB port         (mic in, processed)
MAX98357A  Vin → Pi 5V,  GND → Pi GND
           DIN → GPIO21,  BCLK → GPIO18,  LRCLK → GPIO19
Speaker  → MAX98357A screw terminals (4 Ω)
WS2812   DIN → GPIO10 (SPI MOSI) via 330 Ω,  5V → Pi 5V,  GND → Pi GND
Fan      → GPIO 5V / GND  (or a fan header)
Power    → USB-C 5.1V/3A

4b. AEC-correct wiring (matches Alexa "barge-in", recommended)

Route playback through the XVF3800 so its DSP has the echo reference:

Pi 4 ── USB ──▶ XVF3800  (USB soundcard: mic IN + speaker-reference OUT)
XVF3800 I2S OUT ──▶ MAX98357A ──▶ speaker

Now the array cancels your own music, so "Alexa, stop" works at volume. Zero extra parts —

just use the XVF3800's I2S output instead of the Pi's. (See §7 for why this matters.)

5. Software — step by step (self-contained build)

Flash OS. Raspberry Pi Imager → Raspberry Pi OS (64-bit, Lite). Set hostname

openhome, enable SSH + Wi-Fi. Must be 64-bit (wake word/VAD won't run on 32-bit).

Enable the I2S DAC. In /boot/firmware/config.txt:

```

dtparam=audio=off

dtoverlay=max98357a # or hifiberry-dac

dtparam=spi=on # for WS2812 over SPI

```

Reboot; confirm with aplay -l (playback) and arecord -L (XVF3800 shows as a capture device).

Install Home Assistant (self-contained option): HA Container via Docker, or HA OS on a

second box for option B. Docs: https://www.home-assistant.io/installation/

Wire up Assist (STT + TTS + wake word) — all local:

STT: wyoming-faster-whisper (base-int8 on Pi 4) — https://github.com/rhasspy/wyoming-faster-whisper
TTS: wyoming-piper (voice e.g. en_US-amy-medium) — https://github.com/rhasspy/wyoming-piper
Wake: wyoming-openwakeword, preload model alexa — https://github.com/rhasspy/wyoming-openwakeword
In HA: Settings → Voice assistants → Assist → pick Whisper + Piper + openWakeWord.

Run the satellite on the Pi:

Modern: linux-voice-assistant (ESPHome protocol; media player, timers, LED/button

WebSocket API) — https://github.com/OHF-Voice/linux-voice-assistant

Classic/simple: wyoming-satellite tutorial — https://github.com/rhasspy/wyoming-satellite
Prebuilt Pi image (fastest): PiCompose — https://github.com/florian-asche/PiCompose

Point wake word at alexa and set --mic-command to the XVF3800 capture device

(arecord -D plughw:CARD=...). It now answers to "Alexa."

LEDs. Drive the ring from the satellite's peripheral API (linux-voice-assistant WebSocket)

or a small Python service using adafruit-circuitpython-neopixel-spi (SPI, GPIO10) — see §6.

Music. Add Music Assistant (https://www.music-assistant.io/) → play to the satellite

media player (Spotify/local/DLNA).

(Optional) smarter brain. Assist → Conversation agent → wire an LLM (Google Gemini,

OpenAI, or a local model via Ollama) for open-ended questions Alexa would punt on.

6. Light-ring language (WS2812 ×16, over SPI)

State	Animation	Color
Idle	off / faint breathing	dim white 5%
Listening (post-wake)	full ring on	mint `#3ddc97`
Thinking (STT/intent)	spinner	mint → cyan
Speaking (TTS)	soft pulse	mint
Muted (HW mic switch)	solid	red `#ff5a5a`
Error / no network	double blink	amber `#f0b429`

Drive over SPI (GPIO10), not PWM — PWM's GPIO18 is taken by I2S BCLK (see audit).

adafruit-circuitpython-neopixel-spi needs no root and won't fight the audio clock.

7. Debug & design audit (the parts that bite people)

GPIO clash: LEDs vs I2S. The common rpi_ws281x PWM path uses GPIO18 — which I2S

BCLK also uses. Two fixes, one chosen: drive WS2812 over SPI (GPIO10). Enabled with

dtparam=spi=on. Adjustment: BOM/wiring updated to SPI, not PWM.

Barge-in / AEC reference. If audio goes out the Pi's I2S but the mic array is on USB, the

XVF3800 never sees what you're playing, so it can't cancel it — "Alexa, stop" fails over loud

music. Adjustment: recommend wiring 4b (playback through XVF3800) for true AEC. MVP wiring 4a

still works quietly; barge-in just degrades at high volume.

Power budget. Pi 4 (~15 W) + XVF3800 (~2.5 W) + amp/speaker peaks (~3 W) + 16 LEDs

(up to ~0.9 A at full white). A 5.1 V/3 A supply is marginal if everything peaks. Adjustments:

cap LED brightness (≤30%) and never full-white all 16; for headroom use a 5 V/4–5 A supply or a

small separate 5 V buck for the LED ring + amp. Don't pull >1.5 A through the Pi's 5V pin.

Whisper latency/heat on Pi 4. whisper-base-int8 ≈ 1.5–3 s and warms the SoC. Adjustments:

heatsink + fan (in BOM + enclosure vents); if too slow, move STT to option-B server.

Enclosure acoustics. The speaker can vibrate the mic array and self-trigger. *Adjustments in

SCAD:* mic array mounts on the top plate, speaker chamber sealed below the grille, add a

foam gasket / rubber foot to decouple; keep the LED ring diffuser from rattling.

Onboard-audio pop / wrong device. dtparam=audio=off kills the headphone-jack pop; verify

arecord -L and set the exact XVF3800 plughw: string in the satellite mic command.

32-bit trap. VAD/wake word need 64-bit OS. Flash the 64-bit Lite image.
Wake-word false accepts. openWakeWord alexa threshold defaults can over-trigger near a TV.

Tune threshold / use microWakeWord for a lighter, tighter model.

8. Adjustments made vs the first mockup

Mic upgraded ReSpeaker v2.0 → XVF3800 (newer, cheaper $61, dual USB/I2S, better far-field).
Satellite software wyoming-satellite → linux-voice-assistant (maintained; media/timers/LED API).
LEDs moved PWM → SPI to free the I2S clock (audit #1).
Added the AEC-correct audio path so barge-in matches Alexa (audit #2).
Power supply guidance upgraded for LED+amp headroom (audit #3).

9. Upgrade path

Better ears/voice: XVF3800 "with case" option, or add a second Pi as a dedicated HA/STT server.
Better sound: 2.5" full-range + tuned passive radiator; stereo = second MAX98357A (SD pin = L/R).
Physical mute: hardware slide switch cutting XVF3800 mic power → drives red ring (real Alexa parity).
Matrix/voice PE parity: compare against Home Assistant Voice Preview Edition ($59 prebuilt) —

our build trades polish for a bigger speaker, bigger array, and a hackable brain.