Poor wake word pickup

[vhamburger]

Hey, got my device last week and I notice that it really is poorly performing when using the wake word.
I have set it to "Okay Nabu" and it sits in my living room next to my Alexa (which I want to get rid of...).
However, when I calmly say "Okay Nabu" it never picks up my voice - from like 2m away. I have to raise my voice clearly to let it pick up.
On the funny side - I have a self built voice assistant based on a raspberry pi and mic hat - which picks up my command every time - but it is sitting some 12 meters away in my home office. That cant be correct.

Is there a way to adopt the microphone gain / or set the xmos different to make it work better? Because right now, wife approval factor is below 0 since it never picks up any command.
I am happy to adopt it in ESP home and change the params manually if necessary - just want to make it work reliably.

[RocketRammer]

+1 for wife approval factor. I understand that these are Preview kits, but that's not how consumer law works when you've marketed it as having X feature. Having to sit with my face pressed against it is not how it's marketed.

[Warren-0]

I have this issue too. Okay Nabu will only wait it 1 in 43 times or so. I tried Hey Jarvis and I have to be few inches away from it to respond.

Australian accent here though. I do have the pi zero* and mic hat and that picks up every time. I can use the PE this in the office for now, but not in a room where it must sit 1 to 8 meters away at any time. Just feedback nothing more from me.

[RassK]

Not sure if this is mic / hw issue. Faster whisper with large-v3 model seems very good even with a very rare accent (estonian). In the other hand, wake word works very poorly here too.

[grigi]

I found that if I deliberately say the wake words in a higher pitch it's significantly more likely to get it, but it's really not good enough for me.
I tried to setup openwakeword (the add-on) but i can't get it added to wyoming, nor can I find where to configure it in the assist config. If it was deprecated the documentation isn't making it clear. A bit frustrating.

[jf-uu]

@grigi the Voice PE uses microWakeWord to perform on-device wake word detection, so you won't be able to use openWakeWord with it.

[vhamburger]

Is anyone working on this one or is this only my device. Since ESPHome 2025.2.0 it has gotten worse. I practically have to yell at the device standing 1m in front of it. Is my device broken? Do I have to return it - or is this just "how it is"?
Can anyone let me know if you see a similar pattern?
For comparison - I have built an assist satellite on an RPI with a 2mic hat and it works flawlessly any time - even from across the room. However, it streams the wake word to HA which probably has a more powerful detection engine and noise cancellation capabilities.

The current state makes the Voice PE unusable in my household, no one wants to use it as it is never picking up the wake word. Sad.

[RocketRammer]

I've entirely stopped using mine for Voice @vhamburger and purely use them as speakers now. Raised a few times across issues here that they are terrible at this, but much like the PE, it seems to fall on deaf ears.

For voice now, I've just setup the action button on our Samsung Galaxy Watches to be Assist and it works incredibly well. I use local Whisper/Piper with fallback to OpenAI and it's absolutely Rapid. I wish PE was capable to provide this too.

[RassK]

Same here, unusable without a good wake word support.

[vhamburger]

Hey all - so since no one from Home Assistant or the Voice PE team even got back to this (not sure if just ignoring or too much to do otherwise) I started to experiment around on my own. For the following "fix" you need to adopt your device and slightly edit the firmware yaml - if that is not for you, the home assistant folks may pick this up eventually and might "fix" this in a later release.
I must say I am a bit disappointed that I did not even get a reaction to this - they claim they get so much good response to the device and that it is awesome, which it is don't get me wrong, however, if you look in the issues part in GitHub you can clearly see that I am not the only one with this problem.

However - "Never give up, never surrender" (quote: Captain Peter Quincy Taggart) - so I started to play with some options and I think I found a combination wich works great for my setting (I am a native German speaker - so it might vary for you ;)

I made the following changes to the original yaml:

micro_wake_word:
  id: mww
  models:
    - model: https://github.com/kahrendt/microWakeWord/releases/download/okay_nabu_20241205.4/okay_nabu.json # original: https://github.com/kahrendt/microWakeWord/releases/download/okay_nabu_20241226.3/okay_nabu.json
      id: okay_nabu
      probability_cutoff: 0.78 # Added this to lower detection barrier - may lead to more false positives though

So I switched to an older "okay nabu" model by @kahrendt (thanks for all your hard work on that buddy!) Interestingly, the older model seems to be picking up my voice much easier, better and faster. I no more have to "yell" at the device. The lower probability cutoff helps as well.
Note: I have played with lowering the cutoff only, but that didn't create the wanted results. It helped a bit to detect more wake words, but it also lead to a lot of false positives (mostly during movies where someone said something with "Ok" and a German word that is phonetically similar to "nabu" like "nanu, Kanu, Manu, etc..)

Long story short: Since I switched to an older wake word model it improved a lot for me. Until @kahrendt might improve the existing model, I plan to stick to that one. It also proves that my devices microphones are fine (I could verify that easily since all voice commands always work - it is just the wake word pickups that is flawed).

I hope this might help some people. The only "barrier" to this is that you need to adopt the device in ESP Home. If you have never done it, it is quite straightforward. Just remember you probably need to setup your wifi and api encryption key to make it work - but there are plenty of guides on how to do that.

Enjoy your voice PE. I will keep this updated if I find out more... (or another improvement - still need to test it more with the whole family...)

[StrandmonYellow]

I thought it was just me, but it feels like it got worse indeed with later versions when compared to an erlier one. Is this being picked up? @synesthesiam 🙃

[Warren-0]

I found that if I deliberately say the wake words in a higher pitch it's significantly more likely to get it, but it's really not good enough for me. I tried to setup openwakeword (the add-on) but i can't get it added to wyoming, nor can I find where to configure it in the assist config. If it was deprecated the documentation isn't making it clear. A bit frustrating.

My kids can wake it up using any wake word almost every time, not that I think about it. They are high pitch voices too yes..

[mirml]

Both of my google mini and nest picks wake word far better than voice assistant pe 😔👎

[vhamburger]

Hey - this topic seems to be highly ignored by the Voice PE team - I wonder why that is?
So maybe I do have a faulty device - it is hard to tell. The other day I set up an M5 Stack Mini and it was picking up "OK Nabu" from across the room when Voice PE did not pick it up? How is that even possible with its tiny microphone...

Is there a way to debug the wake word sound? Can you set it to debug mode or something to examine the audio it picks up. I am really upset now since basically my whole family is rejecting the device as it almost never picks up the wake word - never.
Everybody is revering back to Alexa now and is making fun of it... Kind of sad, as I wanted to get rid of the Amazon spy...

[StrandmonYellow]

I don't want to judge because the team is doing many things, but indeed this topic has been open for a while. I have to say the wakeword twice or three times allmost every time for it to trigger correctly. I don't know why that is the case and currently do not have the time to dig deeper into this issue unfortunately.