AsteriskGuru Archives

astcallcenters at wunderk · Guest

Hi all, especially Matt :D

I saw a post from you on here awhile ago about recording outgoing calls
and analyzing them with Sphinx. I am looking to do the same thing. I
was wondering if you could share some of your Sphinx magic with me.

I think I am finally making progress, but it is a steep curve. I am
able to get Sphinx to recognize the number 1, but nothing else from 2-9.
That is using the turtle dictionary. I tried downloading some things
from: http://xvoice.sourceforge.net/xvoice-sphinx/ but now I have to
work the magic to get it all to work together.

Would you be willing to share your dictionary file and other related
Sphinx files required for the recognition, the 'config file' (like this:
http://lists.digium.com/pipermail/asterisk-users/2003-December/030873.html),
and any gotchya's about the related sound stuff - like what file type
you are recording in Asterisk and any sox magic required to get Sphinx
to recognize it?

I can take care of everything else, just the Sphinx part is difficult
for me. Also, how are you keeping track of each file after is analyzed
with Sphinx? Do you need to make a new 'config file' for each file? I
know that it can process a list of sound files to analyze, but if you
have it save the hyp file, it only shows the one file doesn't it? I am
not very far in the process, so maybe I missed something. I just want
to make sure I have a way to keep each file separate after I am able to
actually get the recognition to work.

Thanks for any pointers,

Kevin

astcallcenters at wunderk · Guest

I forgot to add that I'm using Sphinx2 at the moment, but if I had all
the right dictionary files, etc I could probably install a different
version. I don't care what version I use.

Kevin

astmattf at gmail.com · Guest

Hello,

I spent several weeks customizing the dictionaries and script
parameters for the UK automated messages when an invalid call is sent.
The thing is the UK is easy because it's really still one phone
system: BT. The USA is 100 times harder because there is no
standardization on messages or voices used or pace of the message
across the hundreds of phone companies out there.

I didn't get that far on USA digit recognition for those messages
because it took too much processing power to get any degree of
reliability for hundreds of recordings per day unlike with the UK.

What I did was record the ring time as GSM because WAV took up way too
much space. Then I picked out only the recordings I needed to analyze
and normalized and converted them to WAV with the following commands:
# gets the normalization value
/usr/bin/sox -t gsm -b /dir/files -e stat -v
# normalizes the file and converts it to a new WAV file
/usr/bin/sox -t gsm -b /dir/file -w -s -r 16000 -v $normalize_value
file silence 1 500 -45d

Here is the batch sphinx script that I used to kick off the analysis
of the resulting files:
http://astguiclient.sourceforge.net/experimental_code/sphinx2_pltest.pl

I spent way too much time on all of this sphinx stuff and I really
don't look forward to getting back into it as it was a very
frustrating and extremely time-consuming experience.

Good Luck, and let me know how it goes for you :)

MATT---

On 1/16/06, Kevin Bockman <astcallcenters@wunderkin.com> wrote: