How often are new wake word models released for Voice PE?

[formal patrol]

As per the subject really, how often are contributions people have made via https://www.home-assistant.io/voice_control/contribute-voice/ actually upgraded on our Voice PE (presumably via a firmware update) ?

[kind vale]

So far there were 2 iterations, AFAIK. Probably next one will be released soon

[formal patrol]

OK, so no fixed schedule as such like HA releases

[sand hawk]

It's kinda dependent on more samples coming in, or if they find a new/better training technique while researching.

[gentle lodge]

Is it possible to use custom wake words with home assistant voice PE or is it still only possible to use the 3 default options?

[sand hawk]

You'd have to adopt the device in ESPhome and manually manage it to modify it to use a custom wakeword at the moment, and it has to be a MicroWakeWord trained wakeword. Might be possible to make it use OpenWakeWord instead, but they are two different model types.

[gentle lodge]

So its not possible to use a tflite file? Could you explain why this is possible with the 13$ voice assistant but not with the 50$ Voice PE?

[sand hawk]

It's not that it isn't possible, it's that it isn't in the base design, you'd have to modify the firmware to utilize OpenWakeWord (if you are using an OpenWakeWord trained tflite file) or MWW if you are using that. the tflite files are different depending on which technology they are using.

[gentle lodge]

Thanks for the explanation. Is it likely that in a future update of the Voice PE this feature will be implemented?

[sand hawk]

Hard to say, that's a question for the HA/ESPHome folks 🙂

[winter forge]

I'm hoping for this feature eventually... I have yet to talk to someone that's okay with "Okay Nabu" or Hey Mycroft, or Hey Jarvis. All three have been a dealbreaker. I'm not even sure what I'd train it on, but it's not those.

[wicked thunder]

Well, I would be fine with „Hey Jarvis“ if only the performance would be better… What would be fantastic though was the possibility to add your own samples to the training set and just hit the “retrain” button:-)

[gentle lodge]

I tried to use my tflite but unfortunately without success. If I set the tensor_arena_size in the wake_word.json to 8000 it runs stable without crashing every few seconds but the wake word is never detected.

[kind vale]

tflite for MWW is different from OWW, they're not interchangeable.

[gentle lodge]

I don't know a lot about the topic unfortunately. I tried to learn about it a bit today but it seems very complex. I really hope it will be easier in the future to use custom wake words with the voice pe

[next coyote]

Is there any way to do custom wakewords the same way as with the atom echo? Is there an issue with the old setup of streaming the audio and letting the HA server handle wakeword detection rather than on device? Otherwise I'd like to see that as an option

[gentle lodge]

It should be possible to do it if you tinker with the software of the voice PE. But from my understanding you need a different model so it can run on the hardware

[kind vale]

That thing is Openwakeword. It is done on HA side, so satellite is streaming audio continuously. Wake words for OWW can be trained relatively easy.
PE is using Microwakeword instead. It works on device and doesn't stream audio to HA. But wake words are hard to train.

You can take control over PE firmware and make it use OWW. The disadvantage will be constant audio streaming from satellite to HA, while it's waiting for wake word.

[next coyote]

Yeah, I guess what I mean is that if streaming the audio streaming isn't an issue for performance, I do wish that would become an eventual native UI setting, as I do like getting to use custom wake words.

I wasn't sure if the "design" of the atom echo style architecture of streaming audio to the HA instance was flawed or performance intensive or something, which is why the voice PE does not support it

[kind vale]

I will be implementing this in my Koala firmware. Basically how it was with old satellites, where there was drop-down for mww/oww.

[next coyote]

Oh that'll be great ^^
I hope it also someday reaches the official firmware