I have some experience with S4 and now I am switching to PS.
I have discovored some strange things. I am testing some acoustic models (8kHz
telephone) for the PS models. When I used the wave2feat extractor I achieved
97% ACC using the default Training setup (1000 senos, 8 gaussians).
But when I have used the previously developed frontend the ACC droped to 87 %
(used for training and testing).
I was able to reapair the bad ACC by modyfying the language weight paramter
from 6.5 to 2.1. Then the ACC raised again to 96 %. Now using this setup and
the wave2feat features i get almost 98.8 % acc.
Why does the -lw parameter influence the result so much, as I am using a JSGF
grammar for testing isolated words.
The other thing I have noticed is that when I set cmn from none to current the
ACC drops to 95 %. And I get the same results for PS witth and withous CMN.
The result stays 95 % for current and for nose for acoustic modles build with
cmn=current.
Should I use the PS feature extractor and does CMN current work for live
operations?
Man thank you for any answer.
Best regards,
Mirko
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
And with S4 I noticed around 2-3 % absolute improvement. But here the ACC
drops. But what is more strange is that I get the same ACC with ot without CMN
enabled in pocket sphinx. Normaly I think I should get a very bad ACC when the
features don't match.
Thanks Nickolay
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Sorry for late reply. Unfortunately it's hard to say something meaningful
except "it shoudln't be so". Can you try older pocketsphinx version in order
to check if it's a regression for example? If you'll find the change that
caused that it would be helpful. We also need to setup some meaningful test
for JSGF decoding using pocketsphinx. I also noticed that language weight
influences decoding.
. I am using more then 40 h of material.
Why do you train only 1000 senones then?
t what is more strange is that I get the same ACC with ot without CMN
enabled in pocket sphinx. Normaly I think I should get a very bad ACC when the
features don't match
I think pocketsphinx uses CMN configured in model in feat.params. It ignores
your option.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello.
I have some experience with S4 and now I am switching to PS.
I have discovored some strange things. I am testing some acoustic models (8kHz
telephone) for the PS models. When I used the wave2feat extractor I achieved
97% ACC using the default Training setup (1000 senos, 8 gaussians).
But when I have used the previously developed frontend the ACC droped to 87 %
(used for training and testing).
I was able to reapair the bad ACC by modyfying the language weight paramter
from 6.5 to 2.1. Then the ACC raised again to 96 %. Now using this setup and
the wave2feat features i get almost 98.8 % acc.
Why does the -lw parameter influence the result so much, as I am using a JSGF
grammar for testing isolated words.
The other thing I have noticed is that when I set cmn from none to current the
ACC drops to 95 %. And I get the same results for PS witth and withous CMN.
The result stays 95 % for current and for nose for acoustic modles build with
cmn=current.
Should I use the PS feature extractor and does CMN current work for live
operations?
Man thank you for any answer.
Best regards,
Mirko
Hello
Which version of pocketsphinx are you talking about? I believe there were a
regression recently that could cause such problem.
As for CMN, are you sure you have enough data for training the model? It
doesn't look so.
I am using ps 0.6.1 the latest release....
I am using more then 40 h of material.
And with S4 I noticed around 2-3 % absolute improvement. But here the ACC
drops. But what is more strange is that I get the same ACC with ot without CMN
enabled in pocket sphinx. Normaly I think I should get a very bad ACC when the
features don't match.
Thanks Nickolay
Hello Mirko
Sorry for late reply. Unfortunately it's hard to say something meaningful
except "it shoudln't be so". Can you try older pocketsphinx version in order
to check if it's a regression for example? If you'll find the change that
caused that it would be helpful. We also need to setup some meaningful test
for JSGF decoding using pocketsphinx. I also noticed that language weight
influences decoding.
Why do you train only 1000 senones then?
I think pocketsphinx uses CMN configured in model in feat.params. It ignores
your option.