pocketsphinx package on kinetic

asked 2017-06-10 17:53:30 -0500

burtbick
201 ●9 ●13 ●14

I've been working with Indigo on Linux Mint 17.1 but am in the process of upgrading to Mint 18(.1) so I've installed kinetic. I've gotten the motion bits working and am working on the Speech Synthesis and Recognition parts next.

PocketSphinx has been a bit more of a hassle. There was a poacketsphinx ROS package for Indigo but not for Kinetic. So I pulled in the Indigo package and using an example from Gstreamer for Gstreamer 1.0 / Pocket Sphinx I was able to get it working, kind of. It is recognizing words / phrases from my dictionary but it is NO where near as accurate as with the older version on Indigo.

Has anyone done anything with PocketSphinx on Kinetic, and if so how did you improve the accuracy? My reworked recognizer.py file works fine, except for the accuracy. The sample code I based it off of uses a different approach from the original recognizer.py code, so that may be the issue.

Or is there a better speech recognition alternative that you are using with Kinetic and newer versions of ROS?

Thanks,

edit retag flag offensive close merge delete

add a comment

0

answered 2017-06-10 18:58:03 -0500

burtbick
201 ●9 ●13 ●14

I think that I have it figured out, still some more work to do but it is more accurate now. The lm and dict were not getting set so pocketsphinx was using the defaults.

Now using the lm and dict for the RoboCup example accurately recognizes the test phrases. There's some more work to do, but it appears to be working well enough to continue getting things moved over to Kinetic and Linux Mint 18.1

edit flag offensive delete link

Comments

could you please share your new package,

I m working on now same package but I can not manage the solve this problem

Thans

Developer ( 2017-08-08 08:08:35 -0500 )edit

I'll have to blow the cobwebs off. I've been off working on other things since I resolved the issue. Basically I had to find how to pass the Im and dictionary so that they were being properly set. I did some fixes in src/pocketsphinx/nodes/recognizer.py to make it work with the latest code.

burtbick ( 2017-08-08 09:52:19 -0500 )edit

I'll have to go through the code again to refresh, but I think that it was mainly fixing up recognizer.py I''l do some more digging when I get a chance to see if there were any other changes, and what they were.

burtbick ( 2017-08-08 09:54:24 -0500 )edit

thanks if you remember the modifications on recognizer.py data, I want to apply also same ones one my recognizer data

Developer ( 2017-08-08 16:11:21 -0500 )edit

add a comment

0

answered 2017-06-10 20:48:44 -0500

AndyZe

2331 ●38 ●106 ●70

updated 2017-06-11 05:26:37 -0500

gvdhoorn
86574 ●283 ●1432 ●1054 http://cor.tudelft.nl/

There was another question that's almost identical: #q246247.

This is my preferred solution: github.com/UTNuclearRoboticsPublic/pocketsphinx It's accurate and fast until somebody with an accent tries it. ;)

edit flag offensive delete link

Comments

Thanks, I missed searching for the more generic question about an alternative. I'll check out your preferred solution, even though I have pocketsphinx now working in Kinetic.

burtbick ( 2017-06-11 08:26:07 -0500 )edit

Did you get a chance to compare the two? I'm curious if you noticed a performance difference.

AndyZe ( 2017-06-13 12:02:58 -0500 )edit

I got pulled in a different direction for a day or so, but I hope to compare the two this evening and will let you know.

burtbick ( 2017-06-13 14:00:16 -0500 )edit

Did you eventually compare