I've got wyoming / openwakeword. I followed the documentation and added .tflite files to /share/openwakeword. Restarted whole home assistant. No wake words added/made available to the devices. Just the default 4.
Is there somewhere other than on the config page for the device?
You can't choose wake word at voice assistant creation. Only 4 options, none are wake word.
You can't choose added wake words other than the default 4 on the config page for the device.
What type of device are you using?
respeaker lite
I believe that runs microWakeWord on the device (like the Voice PE and others) so openWakeWord won't work. You should check out GitHub - TaterTotterson/microWakeWords: Tater/Home Assistant voice satellite firmware and microWakeWords for VoicePE, Satellite1, and ReSpeaker XVF3800 with microWakeWord, openWakeWord, and nanoWakeWord support. · GitHub
Then why/how have I had openwakeword running under wyoming, and not microwakeword, and yet the device loads the default "ok nabu" and 3 other wake words? It's very confusing. If it's running microwakeword, it shouldn't work, because my HAOS has no microwakeword. If it's running microwakeword, can't I just install the microwakeword addon, and put the .tflite files in a similar folder?
microWakeWord runs on the satellite, not on the host. There is no microWakeWord addon because the wakeword is not running on the centralized server.
I'm confused. I set "ok nabu" as the wakeword from home assistant, not on the satellite. You're saying I didn't, and it's all running on the respeaker? Can I change the wakeword in the .yaml of the respeaker then? Or is its yaml just pointing back to the assistant configuration on the HAOS, which says it should be "ok nabu"?
Older devices used openwakeword. That's a system where the device sends recognition back to a server for processing.
Newer devices use MICROWakeword which uses a model ON DEVICE to process.
So yes you set the model you want on the device FROM HA. But that's just a selector - it is entirely running on the device itself and there is nothing to 'run' on the HA side.
There are threads here for compiling a MWW and installing it in the custom firmware of your device.
Could you link me to a good thread on here or tutorial elsewhere? I'm trodding through tatertotterson github thing, and not making good progress. I have .tflite files.
You're not going to like the answer. Itlf you're already on taters site you're there. The threads point back to where you already are.
The path is not easy and until about this time last year wasn't documented so... The answer is you already probably have it. What you're doing is pretty much bleeding edge.
The whole micro, nano, open wake word world is a real CF right now IMHO. Unfortunate reality of progress of open source projects for local only voice speech/text to speech systems with the ability to be 'always listening for a wake word'. The proprietary devices, Google, Amazon, Apple got this 'pretty right' early on. And now the open source local only systems (pipelines, projects ... ) are working to get the same. Another moving target for the open source projects is the changes in microphone hardware and esp hardware. Long term these are very positive things, however short term now the ESP32-P4 (which will be the sweet spot device) is moving target and the only high quality speaker system, the respeaker based systems is which appears to be only 'semi-open source' needs some competition.
IMHO, the universe of local STT speech (really STLLM is what we want) and TTS needs a new model above micro, nano open word system that does solid local tflite base wake word identification (ideally multiple wake words with different pipelines) using the ESP32-P4 enhanced local AI. Then combine this with a central wake word coordinator that evaluates each speaker and makes a recommendation (or maybe more strongly ... turns off all devices except the strongest) on which device to have the 'conversation/pipeline' on.
Beyond this minor blue sky, my even higher in the thinning atmosphere :
Continuous listening to background sounds, TV, music. And be able to 'remove' this is real time to hear wake words and conversation stream.
Speaker identification by voice print. Probably centrally, so it is easy to add, update and remove speakers.
It is very scary right now, for example Home Assistant enabling all entities by default to voice assistant, and watching device like Home Assistant Voice Preview Edition and other devices using same wake word models wake up to random noises that they think are wake word and then start trying to process the follow on stream of jibberish as voice commands to Home Assistant ... 'Hal start the destruct sequence for my home and do not accept any commands to cancel'
Soap box off...