I have tried about everything I know stop my Wyoming Satellites from switching from idle to listening at random in a silent room. I have tried modifying the wake-refractory-seconds command as well as setting adding the threshold statement. The result is the same whether I use a seeed 2mic hat on Pi2, or bypass it completely using pulse audio with a USB mic and the Pi4 3.5mm line output.
It drives my wife crazy, especially if I am using an AI voice pipeline. ( I have a VPE arriving Monday that will replace my living room wyoming-satellite and appease her).
If someone has else has had this issue I would love to know if it has been solved, or you just live with it.
Not sure the quality of the wakeword matters since the state changes to listening in a silent room. Essentially these are false positives in a silent environment. My wake words were created using the routine posted on Colab [wake word training environment. (Google Colab).
The words themselves respond well.
Try to choose the standard model “ok_nabu” and evaluate the stability of the work.
If the false positives disappear, then it’s all about your model.
I had the same behavior with ok_nabu as well. Perhaps calling them false positives is a misnomer. The devices state changed to listening randomly in a totally silent environment.
However, I believe I may have solved it today. I read about WiFi dropout challenges some have with TP-Link access points, which is what I use. So I re-provisioned a Pi3 today using sayam93’s published instruction that includes using Pulse Audio and Snapcast. I added a Unifi AP-AC-Pro access point to my network and put the newly provisioned Pi3 on it; the
Pi3 has been rock solid for ~10 hours. It leads me to believe that perhaps that connection error reconnects were/are in some way triggering the listening state. In any event the device is working great with Unifi.