So I trained a model for command and control style decoding for the following sentences:
What is your name
Nice to meet you, Rekha
My name is Ravi
Same here
No, I am not
I am from Mumbai
Where are you from?
Which town?
Is it close to Jaipur?
I have to go now.
Great talking to you
I am attaching the acoustic model, grammar and dictionary file for your reference.
The training data is approximately of 2-3 hours and I was trying to recognise a set of 1000 files post the training.
So when I tried to run the recognition on this particular set using this command
Pocketsphinx continuous is not very good at recognizing short files with too high volume. It needs some time to adapt to the level. You need to set -cmninit value in feat.params to something like
-cmninit 71
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
So I trained a model for command and control style decoding for the following sentences:
I am attaching the acoustic model, grammar and dictionary file for your reference.
The training data is approximately of 2-3 hours and I was trying to recognise a set of 1000 files post the training.
So when I tried to run the recognition on this particular set using this command
pocketsphinx_continuous -infile <filename.wav> -hmm /Users/akshaygoyal/Desktop/Culturealley/Sound_08_06_2016_level_1_test/acoustic-0.0 -jsgf conv1.gram -dict conv1.dic -samprate 8000 -logfn /dev/null -kws_threshold 1e-40f -remove_noise yes
it recognises 54% of the utterances as correct. So when I tried running the files which it didn’t recognise using -inmic mode
pocketsphinx_continuous -inmic yes -hmm /Users/akshaygoyal/Desktop/Culturealley/Sound_08_06_2016_level_1_test/acoustic-0.0 -jsgf conv1.gram -dict conv1.dic -samprate 8000 -logfn /dev/null -kws_threshold 1e-40f -remove_noise yes
it recognises it perfectly.
I am also attaching some of the problematic wav files which show this behaviour.
I am unable to understand what is the issue here. Any leads will be appreciated.
Thanks,
Akshay Goyal
Here is the link to the zip having all the files - https://s3.amazonaws.com/language-practice/English-App/Files.zip
Pocketsphinx continuous is not very good at recognizing short files with too high volume. It needs some time to adapt to the level. You need to set -cmninit value in feat.params to something like
Thanks,
This seems to work fine for me.
Out of curiousity what does this parameter do?
ask google to feed your curiosity