Closed KiboOst closed 4 years ago
I'm passing the sensitivity directly into the snowboy Python code, so I'm not sure what happens after that. Does anyone else have any experience with this?
In C source it is a const string.
Then in some python examples it is multiplied by model length sensitivity = [0.5]*len(models)
So depending on model length I guess it isn't limited to 0.1 step, just a float.
In a java example there is:
snowboyDetector.SetSensitivity("0.45"); // Sensitivity for each hotword snowboyDetector.SetAudioGain(2.0); // Audio gain for detection
Also saw in python demo4.py example:
detector = snowboydecoder.HotwordDetector(model, sensitivity=0.38)
I guess we could allow hundredth and run some test. Testing between 4.01 and 4.99 should be sensible. Doesn't seems to be rounded anyway.
Oh, I think I understand the problem. If you installed with a virtual environment, you may not be seeing the latest web UI changes.
Grab the latest rhasspy-web-dist.tar.gz and extract it into your rhasspy
folder. It should overwrite stuff in rhasspy/dist
.
Ahah yes !! Ok snowboy sensitivity is now float. And, I now can see all sentences ini files ! Would it be possible to have this for slots files alos, instead of one merged files ? This make things lot more organized.
Thanks!
Should be possible. Might be easier to use than having to type JSON in directly. The files are already actually split out in the slots
directory.
Yes, all my slots are separate files into /slots folder.
Now I would really be able to use several /intents files to begin testing rhasspy with all my intents so I can really see how it works and if he handle the right intents regarding what I say.
I have updated the issue regarding this multiple intent handling bug, which is not related to having separated intent files as I tried to detail here: https://github.com/synesthesiam/rhasspy/issues/97
Added to TODO list
Hi,
We can set sentivity of snowboy wakeword by 0.1 Dunno if it's limited by a lib or so, but in Snips there was big difference between 0.51 and 0.53 for example. I had fine tuned them a few times before getting a good parameter. It seems that regarding snowboy, I have a similar problem between 0.5 which get too many false positive, and 0.4 which isn't sensible enough.