0
$\begingroup$

Rosanswers logo

Hi,

I'm using pocketsphinx on ROS Hydro to implement speech recognition in my project. Unfortunately, its accuracy is not so good for a few specific words, and this affects system reliability.

Therefore, in order to find a way to improve its efficiency, I've got a few questions:

  1. Is it possible to "train" it? (e.g. I keep saying the words and stating if pocketpshinx understood them right or not, so it gets better with time).

  2. No matter what I say, it always finds the most similar word defined in the .dic file. Is it possible to define a threshold so it'll only output a word if the similarity between the spoken word and the one defined in the file is above this threshold?

  3. Is it ok to manually edit the .dic file in order to include different ways a word can be spoken? Do I need to change anything in the .lm file?

P.S: The speech recognition works really well for almost all the word I defined, but it's really unstable to some specifc ones. For example, the words toilet and lift are usually recognised as the words to the or go to, even though they're very different.

Does anyone have any recommendation of how to improve the accuracy?

Thanks!


Originally posted by gerhenz on ROS Answers with karma: 53 on 2014-08-08

Post score: 0

$\endgroup$

1 Answer 1

0
$\begingroup$

Rosanswers logo

You can adapt the acoustic model of pocketsphinx. But it's not that easy as you described it (even the Sphinx wiki says it's not easy)

Have a look here.

Your other questions I can't answer. I haven't work much with pocketsphix, because it doesn't meet our requirements.


Originally posted by BennyRe with karma: 2949 on 2014-08-08

This answer was ACCEPTED on the original site

Post score: 1


Original comments

Comment by gerhenz on 2014-08-08:
I'll take a look in the link. Thanks!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.