-1

I have implement the speech recognition using PocketSphinx in my android application, everything works fine except the recognizer is too sensitive. It made my application result (speech-to-text) not working properly.

I have already change the threshold until the lowest I can get, which is 1f.

I used English, and limited library with certain grammar, and I make my own dictionary using the Sphinx Knowledge Base Tool.

Even though I used the acoustic model en-us-ptm from the demo of cmuSphinx.

I have done some research that Sphinx have remove_noise function, but I can seems to find it. or is there any other way to make the pocketsphinx recognizer better ?

Thankyou.

Lupint
  • 41
  • 1
  • 8

1 Answers1

0

Latest pocketsphinx has internal noise reduction module. There is no easy way you can remove/reduce noise as they would have done if there is such.

External noise reduction degrades speech recognition accuracy. Make sure you're using updated version of the library and language model

See more about increasing the accuracy of the recognizer: http://cmusphinx.sourceforge.net/wiki/faq#qwhy_my_accuracy_is_poor

Layansan
  • 46
  • 5
  • hi, thankyou for your respond, I used the language model from the demo given, and I tried to understand that link you gave, but wht I understand is that the pocketsphinx in android can not be trained? am I wrong? Thanks again. – Lupint Jan 29 '16 at 11:04
  • Sphinx comes with finest acoustic models but if you want to specialize for small vocabulary application you can perform [Training Acoustic Model](http://cmusphinx.sourceforge.net/wiki/tutorialam) which I don't recommend you to try. The best way is to do some simple acoustic model adaptation to improve speech recognition . Check out this [tutorial](http://cmusphinx.sourceforge.net/wiki/tutorialadapt). – Layansan Feb 01 '16 at 05:03