Speech Recognition Engines

Nested
LMs run on julius
User: lebee_leon
Date: 2/4/2010 2:53 am
Views: 7914
Rating: 6

Hi,
     I have generated arpa format 3-gram LM by CMU SLM Toolkit.


     I  wonder if the document is suit to Julius, for no matter what word inputted to Julius would not succeed to complete the 2nd pass

    On the other hand,  it seems that train-text has to add stance signs <s> </s> to each stance or 1st-pass will leave warning as follow:


WARNING: 00 _default: no tail silence word survived on the last frame, search failed
 
   Are these situations normal? The AM is hmm15.

   Best regards.

--- (Edited on 2/4/2010 2:53 am [GMT-0600] by lebee_leon) ---

Re: LMs run on julius
User: nsh
Date: 2/4/2010 7:02 pm
Views: 2390
Rating: 6

LMs run on julius

> I  wonder if the document is suit to Juliu

Which document?

>  for no matter what word inputted to Julius would not succeed to complete the 2nd pass

Trigram model for 2-nd pass should be reversed, i.e. all words in training text sentenses should go in reversed order

> On the other hand,  it seems that train-text has to add stance signs <s> </s> to each stance or 1st-pass will leave warning as follow:
WARNING: 00 _default: no tail silence word survived on the last frame, search failed

>   Are these situations normal? The AM is hmm15.

You could use any tool to build ARPA lm - HTK, cmuclmtk or SRILM. But to be able to run julius you need to build reversed trigram model correctly

For example see

http://www.voxforge.org/home/forums/message-boards/general-discussion/how-to-create-reverse-3-gram-for-julius3/5

http://julius.sourceforge.jp/forum/viewtopic.php?f=6&t=90

--- (Edited on 2/5/2010 04:02 [GMT+0300] by nsh) ---

PreviousNext