Is the Feature-VAD branch being developed only for the Android Version or will it affect the Desktop version also. I think the current VAD in desktop version passes on only speech segments for further processing.
Feature vad is for both desktop and android. The advantage of feature-vad is that it applies after noise removal, so it's more robust to noise than the current VAD.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
In the current version 0.8,(and also previous versions) of VAD, VAD only gives speech segments to fe_process_raw function, so I am not able to understand what you mean by "presence of silence in utterance", as you have mentioned earlier in this thread.
You have also talked about accuracy dropping drastically. Was this problem always there in pocketsphinx? I have extensively used 0.6, 0.7 versions of pocketsphinx in FSG mode and I never observed any major difference in accuracy.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
In the current version 0.8,(and also previous versions) of VAD, VAD only gives speech segments to fe_process_raw function, so I am not able to understand what you mean by "presence of silence in utterance", as you have mentioned earlier in this thread.
When I wrote "utterance" I mean batch processing because only batch processing deals with utterances. For continuous processing with VAD this was less a problem, but it affected trainign where VAD was not used.
For continuous processing the noise-robustness of the VAD is an issue. Feature-vad is noise-robust, original energy-based VAD is not.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
ok, so seems like the feature-vad branch has been deleted in the svn repo...I'm also looking to get silence within an utterance detected in pocketsphinx. Would appreciate some pointers on this
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
How is feature-vad branch different from the regular pocketsphinx ? Will this be merged into the next PS release ?
Thanks,
Feature-vad implements proper speech processing API that doesn't degrade in presense of silence inside the utterance.
Yes
So you mean the current Pocketsphinx doesn't handle presence of silence inside utterance properly ! Could you elaborate a bit please.
Thanks.
Yes
If there is a lot of silence CMN is not estimated properly and the accuracy drops drastically.
Hi,
Is the Feature-VAD branch being developed only for the Android Version or will it affect the Desktop version also.
I think the current VAD in desktop version passes on only speech segments for further processing.
How will the new Feature VAD be different.
Regards
Pankaj
Feature vad is for both desktop and android. The advantage of feature-vad is that it applies after noise removal, so it's more robust to noise than the current VAD.
Hi Nicole,
In the current version 0.8,(and also previous versions) of VAD, VAD only gives speech segments to fe_process_raw function, so I am not able to understand what you mean by "presence of silence in utterance", as you have mentioned earlier in this thread.
You have also talked about accuracy dropping drastically. Was this problem always there in pocketsphinx? I have extensively used 0.6, 0.7 versions of pocketsphinx in FSG mode and I never observed any major difference in accuracy.
When I wrote "utterance" I mean batch processing because only batch processing deals with utterances. For continuous processing with VAD this was less a problem, but it affected trainign where VAD was not used.
For continuous processing the noise-robustness of the VAD is an issue. Feature-vad is noise-robust, original energy-based VAD is not.
ok, so seems like the feature-vad branch has been deleted in the svn repo...I'm also looking to get silence within an utterance detected in pocketsphinx. Would appreciate some pointers on this
It is merged in trunk
Thanks....