Direct connection from satellite doesn't work

[tannisroot]

If I add microwakeword to Home assistant via wyoming integration, set it as default method and remove these lines from wyoming satellite service on my raspberry pi:

 --wake-uri 'tcp://192.168.1.7:10400' \
 --wake-word-name 'okay_nabu'

(and \ on the line before those)
the wakeword detection works great.
In debug mode, it prints these lines:

DEBUG:root:Loaded models: {<Model.OKAY_NABU: 'okay_nabu'>}
DEBUG:root:Detected Okay Nabu from client 36499818547860

However, if I re-add those lines and make the satellite try to use the wakeword instance directly, be it on the satellite itself (tcp://127.0.0.1:10400) or when it's on another computer (the same one I tested it succesfully first), wakeword detection just doesn't work.
In debug mode, it prints these lines:

DEBUG:root:Client connected: 36619420553939
DEBUG:root:Sent info to client: 36619420553939

and that's it.
Happens both with the docker image and when installed directly.

[jjdenhertog]

@tannisroot How did you configure the wyoming-satellite and wyoming-microwakeword together with the models? I only get to the place where it says "DEBUG:root:Client connected: 36619420553939". But I don't see any messages about the models being loaded.

[tannisroot]

@tannisroot How did you configure the wyoming-satellite and wyoming-microwakeword together with the models? I only get to the place where it says "DEBUG:root:Client connected: 36619420553939". But I don't see any messages about the models being loaded.

I followed the instructions oh how to set it up, then just added it in home assistant via Wyoming integration and changed to it as wakeword method in the voice pipeline I use. Then I removed configuration associated with local wakeword detection, making the satellite fallback to remote (audio streaming) wakeword detection. After that, whenever I said the wakeword it would print about loading the model in debug logs. Although strangely the client would always disconnect after a wakeword detection and the model would be reloaded again, at least according to logs.

[jjdenhertog]

Ah! I actually thought you managed to run the microwakeword locally on a device. I'm trying to setup a wyoming-satellite together with wyoming-microwakeword on a raspberry pi.

That setup works well with open wake word. But it fails on the microwakeword implementation.

[tannisroot]

Ah! I actually thought you managed to run the microwakeword locally on a device. I'm trying to setup a wyoming-satellite together with wyoming-microwakeword on a raspberry pi.

That setup works well with open wake word. But it fails on the microwakeword implementation.

Yes, that's what this issue is about.

[jjdenhertog]

Cool, yes. I'm interested in a solution as well. I would love to test microwakeword on my raspberry pi setup. Openwakeword words alright, but there is still a bit of delay to detect the wakeword. Maybe @synesthesiam can give us a nudge in the right direction 👀

[synesthesiam]

It may be that the satellite isn't sending a "detect" message with the appropriate wake word name. I should add a command line option to set the default wake word model to use if no detect message comes in.

[tannisroot]

It may be that the satellite isn't sending a "detect" message with the appropriate wake word name. I should add a command line option to set the default wake word model to use if no detect message comes in.

In wyoming-satellite, isn't that what --wake-word-name is supposed to do? I did try setting it to --wake-word-name 'okay_nabu' instead of ok_nabu that openwakeword has and it still wouldn't work.