I had a question, not an urgent one but more a case of curiosity. When I'm
using the hub4wsj_sc_8k hmm with my custom lm/dic files in OpenEars and
AllEars, I discovered through trial and error that the first recognition
tended to be more accurate if I overrode the default cmninit value that is in
the hmm (it's 8) and set it during initialization to the most common
approximate cmn return value, which tends to be around 40.
But, it occurs to me that using a hard-coded value here could be a mistake,
since I don't actually know the reason that the cmn is returned as this or
that value, and if it has something to do with the hardware, it isn't future-
proofed since I have no idea what kind of hardware is down the line.
So, my questions: do you have any advice for a way to dynamically derive the
ideal -cmninit value for a given piece of hardware (if it has anything to do
with hardware), generally what influences the cmn that is returned after
recognition, and do you have a recommendation for whether I should be running
-cmn as current, prior or none (currently I'm leaving it as the default of
current)?
Thanks!
-Halle
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
But, it occurs to me that using a hard-coded value here could be a mistake,
since I don't actually know the reason that the cmn is returned as this or
that value
This is actually a good value for feature extraction type used in this model
(-transform dct2)
you have any advice for a way to dynamically derive the ideal -cmninit value
for a given piece of hardware (if it has anything to do with hardware)
CMN value can be different from each model and depends on the way model is
trained. In recent versions CMN will be set automatically from the model. It's
specified in model in feat.params:
-cmninit 56,-3,1
You don't nee to do anything, this value should be plugged in automatically.
And do you have a recommendation for whether I should be running -cmn as
current, prior or none.
Again, it's specified in feat.params
-cmn current
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi Nickolay,
I had a question, not an urgent one but more a case of curiosity. When I'm
using the hub4wsj_sc_8k hmm with my custom lm/dic files in OpenEars and
AllEars, I discovered through trial and error that the first recognition
tended to be more accurate if I overrode the default cmninit value that is in
the hmm (it's 8) and set it during initialization to the most common
approximate cmn return value, which tends to be around 40.
But, it occurs to me that using a hard-coded value here could be a mistake,
since I don't actually know the reason that the cmn is returned as this or
that value, and if it has something to do with the hardware, it isn't future-
proofed since I have no idea what kind of hardware is down the line.
So, my questions: do you have any advice for a way to dynamically derive the
ideal -cmninit value for a given piece of hardware (if it has anything to do
with hardware), generally what influences the cmn that is returned after
recognition, and do you have a recommendation for whether I should be running
-cmn as current, prior or none (currently I'm leaving it as the default of
current)?
Thanks!
-Halle
Hi
This is actually a good value for feature extraction type used in this model
(-transform dct2)
CMN value can be different from each model and depends on the way model is
trained. In recent versions CMN will be set automatically from the model. It's
specified in model in feat.params:
Again, it's specified in feat.params
Cool, thank you for the rundown.