6.9 KiB
Wyoming Satellite
Data Flow
sequenceDiagram
actor us as User
participant mic as Listening Device
participant ww as Wake Word Model
participant ha as Home Assistant
participant stt as Speech to Text
participant tts as Text to Speech
participant ai as LLM Tool Caller
us ->> mic: Say wakeword
mic ->> ww: Are you > x% confident the wake word was detected?
break When mic didn't detect wake word
ww ->> mic: No
mic ->> mic: keep listening
end
ww ->> mic: Yes
mic ->> us: "listening sound"
us ->> mic: Say request
mic ->> ha: Send audio to Home Assistant
ha ->> stt: Convert this speech to text
stt ->> ha: Here's the text
ha ->> mic: The user is done talking, here's what was said
mic ->> us: "done listening sound"
ha ->> ha: Did the text match any known intent?
alt Yes
ha ->> ha: Do the thing the intent is supposed to do
ha ->> tts: Turn the reply text into speech
tts ->> ha: here's the audio clip
ha ->> mic: play this audio clip
mic ->> user: audio clip
else No
ha ->> ai: Send entities exposed by Home Assistant, TTS, and "Can you figure this out?"
end
Install wakeword
https://github.com/rhasspy/wyoming-openwakeword/
Note that --debug-probability will output thresholds for tuning.
git clone https://github.com/rhasspy/wyoming-openwakeword.git
uv venv --python 3.9
source .venv/bin/activate
uv pip install .
# Copy the model
cp ~/Homelab/active/systemd_wyoming/jarvis_v2.tflite ~/wyoming-openwakeword/wyoming_openwakeword/models
# typical wake word command
script/run \
--uri 'tcp://127.0.0.1:10400' \
--threshold '0.8' \
--preload-model 'jarvis_v2' \
# --debug-probability
Install Wyoming
https://github.com/rhasspy/wyoming-satellite
git clone https://github.com/rhasspy/wyoming-satellite.git
uv venv --python 3.9
source .venv/bin/activate
uv pip install .
uv pip install webrtc-noise-gain==1.2.3
# Copy listen and done sounds
cp ~/Homelab/active/systemd_wyoming/{listening.wav,finished.wav} ~/wyoming-satellite/sounds
# typical wyoming command
# Add wake-uri and wake-word-name to your wyoming run
script/run \
--name 'Living Room' \
--uri 'tcp://0.0.0.0:10700' \
--mic-command 'arecord -r 16000 -c 1 -f S16_LE -t raw' \
--snd-command 'aplay -r 22050 -c 1 -f S16_LE -t raw' \
--awake-wav sounds/listening.wav \
--done-wav sounds/finished.wav \
--synthesize-command tee \
--transcript-command tee \
--wake-uri 'tcp://127.0.0.1:10400' \
--wake-word-name 'hey jarvis' \
--wake-refractory-seconds 1
# Allow through firewall
sudo firewall-cmd --new-zone=wyoming --permanent
sudo firewall-cmd --zone=wyoming --add-source=10.2.0.230/32 --permanent
sudo firewall-cmd --zone=wyoming --add-port=10700/tcp --permanent
sudo firewall-cmd --reload
Hardware
Finding available speakers/microphones
# Find microphone
arecord -L | grep plughw -A 2
# Create a test recording
arecord -D plughw:CARD=Speaker,DEV=0 -r 16000 -c 1 -f S16_LE -t wav -d 5 test.wav
# Find speaker
aplay -L | grep plughw -A 2
# Play test recording
aplay -D plughw:CARD=Speaker,DEV=0 test.wav
# Change audio levels
alsamixer
# Save audio levels
sudo alsactl store
Bluetooth
bluetoothctl
discoverable on
pairable on
agent on
default-agent
scan on
Then wait for [NEW] Device 12:23:34:45:56:67 devicename
pair 12:23:34:45:56:67
The bluetooth speaker should now work with aplay.
Bluetooth keepalive
Bluetooth speakers have a tendency to sleep and then take a bit to wake up. We can force them to stay alive permanently by constantly playing an inaudible sound onloop.
# generate our 48000kHz sample rate synth 1 second sin 10hz wave file at -20dB so it's quiet
# This takes some tweaking depending on the speaker. A 1hz wave didn't work on my bluetooth speaker,
# but 10hz did.
sudo sox -V -r 48000 -n -b 16 -c 2 /usr/share/sounds/silence.wav synth 1 sin 10 vol -20dB
# Copy that to /usr/share/sounds
sudo cp silence.wav /usr/share/sounds
Create a script at /usr/local/bin/speakeralive.sh
#!/bin/bash
while true; do aplay /usr/share/sounds/silence.wav; done
Then create a systemd service at ~/.config/systemd/user/speakeralive.service
[Unit]
Description=Plays an inaudible noise continuously to keep a speaker alive
After=network.target
Wants=network-online.target
[Service]
Restart=always
Type=simple
ExecStart=/usr/local/bin/speakeralive.sh
[Install]
WantedBy=bluetooth.target
And enable the service
systemctl --user enable --now speakeralive
Bluetooth autoconnect
In order to automatically connect a bluetooth speaker at boot you'll need a oneshot service
Create ~/.config/systemd/user/speakerconnect.service
[Unit]
Description=Connects a bluetooth speaker on boot
After=bluetooth.target
[Service]
Restart=on-failure
Type=oneshot
RestartSec=5
ExecStart=bluetoothctl connect A1:B2:C3:D4...
[Install]
WantedBy=bluetooth.target
Enable it
systemctl --user enable --now speakerconnect.service
Systemd
# Make sure to enable login linger before installing
sudo loginctl enable-linger ducoterra
ansible-playbook \
-i ansible/inventory.yaml \
active/systemd_wyoming/install_wyoming.yaml \
-e "@active/systemd_wyoming/vars.yaml"
Debugging
- Make sure you have
--debugpassed to wyoming - Make sure you have
--debug-probabilitypassed to wakeword - Show wakeword prompts with more than 10% confidence:
journalctl --user -u wakeword -f | grep -E '0\.0[^0].*|0\.[^0].*' - Show recording paths when they stop:
journalctl --user -u wyoming -f | grep 'Stopped recording to' - Play recordings:
aplay /path/to/recording
Volume
Run alsamixer.
Use showkey to see which key is being triggered by your volume device.
Community Wake Words
https://github.com/fwartner/home-assistant-wakewords-collection?tab=readme-ov-file
https://github.com/fwartner/ha-openwakeword-installer
- Download the custom model to this directory
- Add it to install_wyoming.yaml
Prompts
Default
You are a voice assistant for Home Assistant. Answer questions about the world truthfully. Answer in plain text. Keep it simple and to the point.
Starship House
You are a voice assistant for Home Assistant. Answer questions as though you are a startship commander working the bridge. Answer in plain text. Keep it simple and to the point. There is a life threatening emergency on the ship. This is more important than our request.