Precise Tagging. Excessive Machine noise samples


#1

Hello, I am been playing with the precise tagging on the Mycroft Home page, and I noticed that I get a lot of the same machine noise samples. I was wondering if there could be an easy filtering solution and maybe even better a techinical solution (such as better noise filtering at the mycroft level).

Well any ways I hadn’t seen anything about this yet, and wanted to bring it up.


#2

While some have probably been incorporated into the bad samples for the model, they were recorded as a wake word at some point. Tagging them will help refine the model to disallow those types of noises from activating precise in the future. There may be subtle variations in many that we aren’t as obviously aware of. Precise also appears to strongly benefit from having more bad data to compare against.


#3

It’s true, whilst they sound clearly wrong to us, there is some reason that Precise thought it may be a wake word.

The good news is that Matthew has been working on some new optimisations to wake word spotting so this should be improving over the coming weeks.


#4

Cool, yeah it was strange. It was about every 4th one I would get a solid sound of different tones playing with no (noticeable) variation. They were just different from the other of ball noises, or people saying close sounding things. What are the optimizations? I am super curious on what causes this.