This is the problem I faced using pocketsphinx_continuous -infile test.wav
I have two sentences in the test file. When I tried the pocket sphinx on the
file, the second sentence was recognized correctly, but the first one was
complete mess!
I opened my wav file in audio editor and copy/paste the first sentence after
the second one (so 3 sentences, where first and third was completely same) and
tried again,
this time the result for the third sentence was much better!
So, i'm confused! Why it's returning two different results for the same
sentence?
Am I doing anything wrong?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
pocketsphinx in live mode uses prior channel parameter estimation which called
CMN estimation which improves over time. You can see it in the log. You can
try to specify initial CMN value with -cmninit to make it converge faster.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Batch mode uses different CMN (current) which means the values are normalized
over the utterance. For short utterances it's worse, for long it's better.
Best approach would be to use current in the beginning and then fallback to
prior but this thing is not implemented.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi
This is the problem I faced using pocketsphinx_continuous -infile test.wav
I have two sentences in the test file. When I tried the pocket sphinx on the
file, the second sentence was recognized correctly, but the first one was
complete mess!
I opened my wav file in audio editor and copy/paste the first sentence after
the second one (so 3 sentences, where first and third was completely same) and
tried again,
this time the result for the third sentence was much better!
So, i'm confused! Why it's returning two different results for the same
sentence?
Am I doing anything wrong?
pocketsphinx in live mode uses prior channel parameter estimation which called
CMN estimation which improves over time. You can see it in the log. You can
try to specify initial CMN value with -cmninit to make it converge faster.
What if I use the batch mode? (pocketsphinx_batch)
Is that the same? Or that will be more accurate from start?
Hello
Batch mode uses different CMN (current) which means the values are normalized
over the utterance. For short utterances it's worse, for long it's better.
Best approach would be to use current in the beginning and then fallback to
prior but this thing is not implemented.