Speakerphone Assist
Use a USB conference speakerphone as a Home Assistant voice satellite — say a wake word, give a command,
hear the answer. The official Linux satellite software, connected and tuned the right way, with the parts
that are hard to find written down: Proxmox LXC audio, PipeWire without udev, and your own wake word.
🌐 English · Lietuvių · Русский
ℹ️ Made with AI: the scripts and texts were written by Claude (Anthropic) together with Robertas and tested on
real hardware; the illustration was created with AI.
Tested on 2026-10-10 with an EMEET Luna Plus (A350 USB dongle) in an unprivileged Proxmox LXC container:
"Okay Nabu" → chime → «Который час?» ("What time is it?", asked in Russian) → the answer is spoken by the speakerphone.
Wake word detected 3 times out of 3. Home Assistant found the device on its own.
Why a conference speakerphone
A speakerphone is made for exactly this job: a microphone array that hears a whole room, echo cancellation in
hardware, a loud speaker and a single USB plug (or a small USB radio dongle). You buy it ready-made — no
soldering, no microphone boards, no 3D-printed case.
How it works
speakerphone ──USB──► Linux (PipeWire) ──► Linux-Voice-Assistant ──ESPHome protocol, port 6053──► Home Assistant
▲ (speech-to-text → intent → text-to-speech)
└──────────────────────────────── spoken answer ◄─────────────────────────────────────────────┘
- Linux-Voice-Assistant (LVA) is the official satellite program from the Open Home Foundation, still marked
experimental by its authors
(OHF-Voice/linux-voice-assistant). It listens for the wake
word on the satellite itself and talks to Home Assistant over the ESPHome protocol.
The older
wyoming-satellite is deprecated — do not start new installs with it.
- Home Assistant does the rest with your Assist pipeline (speech-to-text, conversation, text-to-speech).
Choose your path
| Your setup | What to do |
|---|
| Speakerphone plugged into the Home Assistant OS machine | Use the official Assist Satellite app from OHF-Voice/apps. Not tested by us. |
| A normal Linux machine (mini PC, VM, 64-bit Raspberry Pi) | PipeWire finds the speakerphone by itself there — follow the official install guide, then tune it as below. Our setup-lva.sh is made and tested for containers. |
| Proxmox with an LXC container | Steps 1–4 below. ✅ This is what we tested. |
Before you start
- EMEET only: name the speakerphone and update its firmware with the EMEETLINK app before you plug the
dongle into the server. The app exists only for Windows and macOS and works over the USB cable.
- Find your card id. Plug the speakerphone into the Linux machine and run
arecord -l:
card 2: A350 [EMEET Adapter A350], device 0: USB Audio → the card id is A350.
- One owner per card. The speakerphone must not be used by anything else (the host, another container,
another program). If something holds it, PipeWire will not start.
Step by step (Proxmox LXC)
1. Create a container — Ubuntu 24.04 template, unprivileged, nesting on, 2 cores, 2 GB RAM, 8 GB disk.
Leave it stopped.
2. On the Proxmox host, give the container the sound devices:
./pve-speakerphone-lxc.sh 120 # dry run: lists your sound cards and shows what will be added
./pve-speakerphone-lxc.sh 120 --apply # writes it (backs up the old config first)
It adds three things to /etc/pve/lxc/120.conf — access to ALSA devices (c 116:*), a bind mount of /dev/snd
and an idmap for the audio group — and the matching root:29:1 line to /etc/subgid. Without the idmap an
unprivileged container sees the devices as nobody:nogroup and cannot open them.
3. Inside the container:
./setup-lva.sh --card A350 --name living-room
It installs PipeWire and LVA, creates the user lva, describes the speakerphone to PipeWire statically
(see Troubleshooting — containers have no udev) and starts the linux-voice-assistant service.
Options: --wake hey_jarvis (wake word), --user, --ref (LVA version).
4. In Home Assistant: Settings → Devices & services → Discovered → lva-‹MAC address› (ESPHome) → Add.
After adding, the device carries the name you gave with --name. If nothing is discovered:
Add integration → ESPHome → the container's IP, port 6053.
Tune it in Home Assistant
Everything is on the device page — no files to edit.
| Setting | What it does | Our advice |
|---|
| Assistant (×2) | Which Assist pipeline answers | Two slots: one wake word can use an English assistant, the other a different language |
| Wake word (×2) | Bundled: Okay Nabu, Hey Jarvis, Alexa, Hey Mycroft, Okay Computer, Hey Home Assistant, Hey Luna, … | "Hey Luna" suits an EMEET Luna 🙂 |
| Wake word sensitivity | The detection threshold (higher = stricter) | Wakes up by itself → raise it. Misses you → lower it a little |
| Mic volume / auto gain / noise suppression | Software processing before the wake word | A speakerphone already cleans the sound — start with gain and suppression off |
| Finished speaking detection | When it decides you have stopped talking | relaxed if it cuts you off, aggressive for faster answers |
| Mute, thinking sound | Microphone off; a sound while waiting | — |
Three things that are not on that page but matter:
- Expose your devices to Assist (Settings → Voice assistants → Expose). Otherwise the satellite understands
you perfectly and Home Assistant answers that it has no access to the device.
- Speech-to-text speed is the bottleneck, not the satellite. In our test a large Whisper model on a CPU took
~7 s per command. A smaller model, a GPU or Speech-to-Phrase for home commands makes it much faster.
For Lithuanian: LT_ausys — about 0.2 s per command on a CPU.
- After the chime, say only the command. Repeating the wake word turns it into the command.
Speaker volume: the media player volume in Home Assistant, and the hardware level with alsamixer -c <card id>
(control PCM). Microphone hardware level: control Mic (85 % on our unit).
Your own wake word ("Okay Jenny", "Alisa", …)
LVA loads every *.json it finds in its wake word folders; each one appears in the Home Assistant wake word list.
setup-lva.sh adds the folder ~lva/wakewords for your own models. Two model types:
{ "type": "openWakeWord", "wake_word": "Okay Jenny", "model": "okay_jenny.tflite",
"openWakeWord": { "probability_cutoff": 0.6 } }
{ "type": "micro", "wake_word": "Okay Jenny", "model": "okay_jenny.tflite",
"micro": { "probability_cutoff": 0.85, "feature_step_size": 10, "sliding_window_size": 5 } }
Put okay_jenny.json and okay_jenny.tflite into /home/lva/wakewords/, run
systemctl restart linux-voice-assistant and choose it in Home Assistant. The model must be .tflite.
openWakeWord training (below) gives you the first kind; the second is for microWakeWord models.
Making the model. The openWakeWord training notebook creates
thousands of synthetic voices saying your phrase and trains a small model (about an hour on a GPU or in Colab).
What we measured while training our own non-English name:
- Check the pronunciation first. Generate a few samples and run them through speech recognition. For a
non-English name, voices of another language may say it better: for "Alisa" the best Russian voice was right 12/12
(other Russian voices far less), English ones 0/20.
- Synthetic voices alone are not enough. Our model recognised synthetic samples 95–98 % of the time, but a
real person through the speakerphone only 43–59 %. Add 50–100 recordings of your own voice made through the
same speakerphone.
- Write the exam before training: for example "at least 9 of 10 real calls, at most 1 false wake a day", and test
on recordings that were not used for training.
Asking an AI assistant for help? Give it AI_CONSULTANT_BRIEF.md — it contains this method.
Troubleshooting
| Symptom | Cause and fix |
|---|
PipeWire shows only a dummy device (auto_null) | Containers have no udev, so PipeWire cannot discover the card. setup-lva.sh adds static nodes (~lva/.config/pipewire/pipewire.conf.d/50-speakerphone.conf). |
| PipeWire does not start | Most often the card is busy — something else has it open. Check journalctl --user -u pipewire. |
Devices inside the container belong to nobody:nogroup | The audio group idmap or the /etc/subgid line is missing — run pve-speakerphone-lxc.sh again. |
arecord/aplay as root inside the container says "No such device" | Not a missing device — container root is not in the audio group. Test as the lva user: runuser -u lva -- arecord -l. |
| "Access to … is not granted" / "not exposed" | Expose the device to Assist (see above). |
Failed to load cookie file | Harmless. |
| The speaker refuses 16 kHz | Normal for EMEET (plays 48–96 kHz). PipeWire resamples. |
| It suddenly "hears" only noise, no USB errors | The EMEET dongle can lock up without disconnecting. Switch the speakerphone off and on. |
Deaf after the dongle dropped off USB (dmesg shows a USB disconnect and reconnect) | The audio server probably still holds the old device. Restart it and the satellite: runuser -u lva -- env XDG_RUNTIME_DIR=/run/user/$(id -u lva) systemctl --user restart pipewire pipewire-pulse wireplumber && systemctl restart linux-voice-assistant. We met this on our own (non-LVA) speakerphone setup; not reproduced with LVA. |
| Two identical speakerphones | Tell them apart by serial number (/dev/snd/by-id/). Card ids (A350, A350_1) may swap after a reboot — one speakerphone per container is the simplest setup. |
Tested with
| |
|---|
| Speakerphone | EMEET Luna Plus with A350 USB dongle (328f:007b): microphone mono 16–48 kHz, speaker stereo 48–96 kHz |
| Host | Proxmox VE 9.2.6, unprivileged LXC, Ubuntu 24.04 |
| Software | PipeWire 1.0.5, WirePlumber 0.4.17, Python 3.12, Linux-Voice-Assistant b49fd8d (2026-09-29) |
| Home Assistant | 2026.10.0 |
| Scripts | pve-speakerphone-lxc.sh and setup-lva.sh run end-to-end on a fresh container — tested with the host's built-in sound card, because the speakerphone was busy at the time; the voice test above went through the same steps by hand (passing only the speakerphone's device nodes instead of all of /dev/snd). setup-lva.sh is safe to run again |
License
The scripts and texts in this repository: GPL-3.0-or-later. Linux-Voice-Assistant itself is not included; the
setup script downloads it from its own repository (Apache-2.0).
