PocketSphinx generally works very well for me, and OK for my wife, though recognition for my wife is much worse for certain keywords. I'm trying to figure out why these keywords aren't triggering for her (and me, for the last test). I've attached some samples, and would greatly appreciate any help.
I explained in the readme, but in short:
For the first test, 4/5 keywords are recognized for me. 0 for her.
For the second test, 5/5 keywords are recognized for me. 0 for her.
For the third test, 0/5 keywords are recognized for both of us.
Wow, I didn't know -beam existed. That should help a lot. Thanks!
For the second test (hide and go seek), any ideas why I need such a high threshold to detect my wife (1e-90)? Unfortunately, it looks like setting it that high generates false positives (it detects 6/5 for me). I think her pronunciation sounds pretty good, and that phrase is too short to split, so I'm not sure what I can do.
I'm also curious about the last test (smile for the photo) since that fails even for me without high thresholds. Most phrases I've tried work great for me, so the low accuracy is unusual -- especially with just 5 syllables. Do you know what's happening there?
Last edit: Brian Nicholson 2017-01-27
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello,
PocketSphinx generally works very well for me, and OK for my wife, though recognition for my wife is much worse for certain keywords. I'm trying to figure out why these keywords aren't triggering for her (and me, for the last test). I've attached some samples, and would greatly appreciate any help.
I explained in the readme, but in short:
For the first test, 4/5 keywords are recognized for me. 0 for her.
For the second test, 5/5 keywords are recognized for me. 0 for her.
For the third test, 0/5 keywords are recognized for both of us.
Samples here
Thank you!
Hello
There are few things
1) Thresholds above 1e-50 do not work without beam configuration. You need to add
-beam 1e-100
to try threshold 1e-802) You need higher threshold for your wife and it is better to split keyphrases on two:
back to the /1e-30/
drawing board /1e-30/
3) Your wife says
drawing
asJH AW IH NG
, you need to use kws file like thisand dictionary like this
Or use higher threshold.
Wow, I didn't know
-beam
existed. That should help a lot. Thanks!For the second test (hide and go seek), any ideas why I need such a high threshold to detect my wife (1e-90)? Unfortunately, it looks like setting it that high generates false positives (it detects 6/5 for me). I think her pronunciation sounds pretty good, and that phrase is too short to split, so I'm not sure what I can do.
I'm also curious about the last test (smile for the photo) since that fails even for me without high thresholds. Most phrases I've tried work great for me, so the low accuracy is unusual -- especially with just 5 syllables. Do you know what's happening there?
Last edit: Brian Nicholson 2017-01-27