If you're on Windows, just use WSL, this is not presently maintained for Windows.
Presently, you need to use Python 3.11, due to piper-tts issues. Install Python AND the Python dev package
Check requirements.txt and look for the other packages you may need.
I suggest that you create a virtual environment to keep the packages needed to make this work separate from the rest of your OS.
# First, change directory into this repo
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
You'll need an MQTT broker to run bloob. The most common seems to be mosquitto
Arch: sudo pacman -S moquitto
Fedora: sudo dnf install mosquitto
Debian/Ubuntu: sudo apt install mosquitto
etc etc, it's mosquitto basically everywhere.
No configuration is necessary for a single-device setup, however I expect to later find some configs that may be useful.
For a multi-device setup, you'll need to make sure that your broke is accessible from other machines on the local network (where the default only permits connections from localhost). To do this with mosquitto, in a default and with no-added-security setup, add
listener 1883 0.0.0.0
allow_anonymous true
to your /etc/mosquitto/mosquitto.conf file.
If mosquitto is launched as a daemon (which seems to happen automatically on Debian, as it did on my Pi), I believe that it uses this configuration file by default, but launching it manually through your terminal will require mosquitto -c /etc/mosquitto/mosquitto.conf to specify that you want to use the config.
Even this may not work, and that's because the system running it may have a firewall running. If you use firewalld, here's how you can add the default port.
sudo firewall-cmd --add-port=1883/tcp --permanent
sudo firewall-cmd --reload
While STT and TTS models can be downloaded automatically, Wakeword / Instant Intent words cannot, presently. Therefore, I suggest you go to the OpenWakeWord Releases and download some models (specifically the ones ending in .tflite), which you'll put into ~/.config/bloob/ww/. This directory is created automatically on first run (but remember, it won't pick up your voice without these models and a restart).
This project needs you to build things yourself. Luckily, Go makes this easy.
Arch: sudo pacman -S go
Fedora: sudo dnf install go
Debian/Ubuntu: sudo apt install golang
Then, from the repo's root, you can just run:
source .venv/bin/activate
mosquitto -d
./run.sh
Which will activate the Python .venv, launch Mosquitto, and build/launch the Orchestrator.
If you'd rather build things yourself, or not build them on every launch, or any other situation where you want to launch things more manually (which, realistically, is most of the time), you can do your venv and MQTT stuff, build the Go things, then just run .src/orchestrator/orchestrator.
All you should need on your system is Golang.
cd src/orchestrator
go build -o orchestrator orchestrator.go mqtt.go utils.go cores.go
cd ../utils/intent_parser
go build -o intent_parser_util_bb_core intent_parser_util.go checks.go mqtt.go utils.go
cd ../../../
You will NOT see all of the logs in the output of the Orchestrator's main.py (just the Orchestrator's logs). I've tried to ensure that this is enough information to broadly understand what's going on, but it may not be enough if you're trying to debug more major issues, so logging is handled over MQTT.
Each Bloob instance will log to bloob/device-id/logs, so subscribing to bloob/+/logs should get you all logs no matter what your device ID is set to. You can use something like mosquitto_sub for this (if debugging on your local machine, the command would be mosquitto_sub -h localhost -t bloob/+/logs), or you could download a GUI app like MQTTX.
Using a large Whisper model causes unusual and significant slowdowns just before transcription.
The default is distil-small (which is not the smallest), and this worked near-instantly for me, however I choose to use Systran/faster-distil-whisper-medium.en due to the higher quality transcription, though this may be unreasonably slow on less powerful machines.
You can use any Whisper model compatible with faster-whisper.