A voice satellite from a ReSpeaker XVF3800

A far-field microphone that actually hears you across a room, wired into Home Assistant’s voice pipeline with ESPHome — and a device file only forty lines long, because a community packages architecture carries everything else. This is the build we run in three rooms: the hardware, the config, the flash, and the lessons that only show up once two of them share a ceiling.

Why this board

The ReSpeaker XVF3800 pairs a four-microphone circular array with an XMOS voice DSP that does echo cancellation, beamforming and noise suppression in hardware, before the audio ever reaches the microcontroller. In practice that is the difference between a satellite that hears the wake word over its own music and one that goes deaf the moment anything plays. The attached XIAO ESP32-S3 runs ESPHome, so the whole thing shows up in Home Assistant as a native voice assistant device — wake word on the device, audio streamed to whatever speech-to-text your pipeline uses, reply played back through a speaker on the board’s amplifier output. Nothing leaves the building.

ReSpeaker XVF3800 USB 4-mic array — the variant with the XIAO ESP32-S3 socketed on the back, sold as the ESPHome/Home Assistant version.

A small speaker — on the board’s speaker terminal, for responses and chimes; any small full-range driver works.

USB-C power — and, for the first flash only, a data-capable cable to the computer running the ESPHome dashboard.

Home Assistant with ESPHome — and a voice pipeline already configured (speech-to-text, a conversation agent, text-to-speech).

The architecture: forty lines over a pinned upstream

Driving the XVF3800 from ESPHome takes real code — custom components for the DSP, the LED ring, timers, the voice-assistant glue. None of it belongs in your config. The formatBCE Respeaker-XVF3800-ESPHome-integration project packages all of it as remote ESPHome packages: your device file names the repository, a list of package files, and a commit hash — and ESPHome assembles the rest at build time. The hash is the discipline that makes this safe: one ref selects the YAML packages, the C++ components and the XMOS DFU firmware together, so every satellite builds reproducibly from the same upstream state, and upstream can refactor all it likes without touching your rooms until you choose to move the pin.

THE WHOLE DEVICE FILE · esphome/respeaker-living-room.yaml

# =============================================================================
# ReSpeaker satellite — device file (formatBCE packages architecture)
# Everything except the settings below comes from the upstream packages.
# =============================================================================

substitutions:
  name: respeaker-living-room
  friendly_name: ReSpeaker Living Room

  # Pin to an exact upstream commit. This one ref selects the YAML packages,
  # the C++ components AND the XMOS DFU firmware together, so every satellite
  # builds reproducibly from this exact upstream state. To take a future
  # upstream update: check the repo's commits page, replace the hash,
  # reflash ONE satellite first and live with it for a day.
  respeaker_ref: d52dd1acc740c1ceeb2c61ce1299341f808c808c

packages:
  respeaker:
    url: https://github.com/formatBCE/Respeaker-XVF3800-ESPHome-integration
    ref: ${respeaker_ref}
    refresh: 1d
    files:
      - packages/base.yaml
      - packages/hardware.yaml
      - packages/voice-assistant.yaml
      - packages/leds.yaml
      - packages/timers-alarm.yaml

# A remote package cannot read your secrets, so credentials live here.
# ESPHome merges these blocks with the packages' own wifi/api/ota blocks.
wifi:
  ssid: !secret wifi_ssid
  password: !secret wifi_password

# api encryption can be added once the build is settled — one change per
# flash. When everything works, add:
#   api:
#     encryption:
#       key: !secret api_key
# (generate a key, add api_key to secrets.yaml, reflash, re-enter in HA)

ota:
  - platform: esphome
    id: ota_esphome
    password: !secret ota_password

AND THE MATCHING ENTRIES · esphome/secrets.yaml

wifi_ssid: "YourNetwork"
wifi_password: "your-wifi-password"
ota_password: "a-long-random-string"

Flash and adopt

1 · Create the device. In the ESPHome dashboard, make a new device, replace its generated config with the file above (own name per room), and add the three secrets.

2 · First flash over USB. Plug the XIAO’s USB-C into the dashboard machine and install via the connected port. The first build compiles the upstream components and can take a while; every flash after this one is over the air.

3 · Adopt in Home Assistant. The ESPHome integration discovers the device; add it, assign it to its room’s area, and in its device page pick the voice assistant pipeline and the wake word. The board’s LED ring confirms wake and listening states out of the box.

4 · Repeat per room. One file per satellite, identical but for name and friendly_name — the pin keeps every room on the same build.

Watch out for

Never float on upstream main. Unpinned, every rebuild silently takes whatever the repository looks like that day — and a refactor upstream once renamed the package files themselves, which would have broken every satellite on its next routine flash. Pin the hash; update deliberately; reflash one unit first.

One change per flash. A satellite that stops answering could be the Wi-Fi block, the api key, the new ref or the wake word — if you changed all four at once. Settle each change before the next; it is also why api encryption is a later step, not a day-one one.

Two satellites, one voice. Where two units can hear the same spot, both wake and the pipeline answers twice. Give the second unit a different wake word or lower its wake sensitivity, and place arrays so their main coverage overlaps as little as the room allows.

Playback on the satellite deafens it less than you fear — but not zero. Hardware echo cancellation handles the satellite’s own speaker well; a separate loud media player beside it is the hard case. An automation that pauses or ducks nearby media on wake-word detect makes the second sentence of every exchange land.

Credit where due: the heavy lifting here is formatBCE’s integration — packages, components and XMOS firmware. This page is the thin, reproducible way to sit on top of it.