Weird wake word false positives.. How to start troubleshooting?

So I’ve been playing with Snips quite a bit over the past couple of weeks, and have run in to a very strange, and very frustrating issue. I am not using a custom wakeword, but have chosen the ‘built-in’ “Jarvis”. While this seems to work fairly well (recognized 8 times out 10), I get some of the craziest false positives. For example, if I use the word “term”, it recognizes it as the wake word. “Term” and “Jarvis” do not sound anything alike. Has anyone else noticed this? There are some other words and phrases that seem to also trigger the wake word detector, such as “use case”, but not as frequently or reliably. But when I speak the word “term”, it is more reliable then “Jarvis”.

Other then that, I have an assistant with a roughly 200 word vocabulary, and it is amazingly accurate beyond the false wake word detection. I am reluctant to define a custom wake word to work around this, as there are many people who would be using the system, with various accents and inflections, and some ESL people as well, so I’d like to stick with a built in wake word that has more data and training associated with it.

FWIW, I am from the United States, Middle Atlantic region, and have no unusual accent. The “term” false trigger happens from anyone I’ve tested with, regardless of accent or inflection.

Any suggestions on where to begin to work to remove this false trigger would be appreciated.
.

It is unfortunately a known fact that other wakewords did not get as much love as the “Hey Snips” one. You can try to play with the sensitivity of the wakeword in /etc/snips.toml. Pas that, there’s not much to do

1 Like

Hello,

By any chance, are customized hotword more accurate ?
I mean are false detection reduced compared to standard ones?