Voice PE + Experimental MWW Model Performance (specifically Okay Computer)

[buoyant orbit]

I have a VPE that I'm using with the MWW "Okay Computer" model but I feel like the recognition is weak.

To get the VPE with this wake word, I had to adopt the VPE in my ESPHome dashboard and write a custom yaml where I set the probability_cutoff 0.05 as follows

micro_wake_word:
  models:
    - model: github://esphome/micro-wake-word-models/models/v2/experiments/okay_computer.json
      probability_cutoff: 0.05

to get the acceptance rate to be anywhere vaguely acceptable, and, even then, I would say that the VPE still rejects "Okay computer" as a wake word about ... 50% of the time (before it was like 99% of the time when I didn't specify probability_cutoff and it used the default 0.97). There's also a false trigger rate with this 0.05 probability_cutoff that is non-zero and kinda annoying

Is anyone running a non-standard custom wake word on VPE and seeing acceptable performance? If so, what's your VPE YAML look like?

[quartz wind]

I trained my own, "Hey Robot" which I use on the VPE with a 98% probability cutoff which I trained using https://github.com/TaterTotterson/microWakeWord-Trainer-AppleSilicon

[buoyant orbit]

I used that repo to train a different "okay computer" wakeword model and it seems to work better with its default provided probability cutoff

[dusky stump]

I think I used the one you trained for mine. Seems to work decently, but I do need to enunciate when using it.