Speech Recognition Engines

Flat
Combining HMM with MLP for phone recognition
User: manju
Date: 7/23/2013 10:06 am
Views: 2756
Rating: 3

Hi

I have built   separate models for HMM (using htk) and Neural nets (MLP using quicknet)  for phone recognition using 35 phones.  Now, i would like to combine both HMM and MLPs.

I m planning to do this way. I will get confidence scores  (or probablities)  from each  HMM  and  MLP  systems and using these, i will do some post processing on it to get improved perforamance.

MLP gives me prosterior probs for each frame. But, how  to  get HMM  confidence score (or probablity)   i.e.    what is the  value of a particular  segment  (or frame)  classification  across  all the  phones???.  

after googling i found out that we need to get lattice files for each test file using HVite and then using lattice_tool of SRILM we can get confidence scores. is this the correct way ??

After this i tried to use Hvite for lattice generation using following command...

 HVite -T 1 -X lat -D -A -n 32 -m -w -q tvaldm -z lat  -H ../train/models//hmm8/macros -H ../train/models//hmm8/hmmdefs -S kk   -o SWT -w wdnet.txt -p -10.0 -s 0 ../intrmed_fils/newDict ../train/models//monophones0

But i am not seeing any *.lat (lattice) files generated. Please tell me what is the problem with it.

Also, is it possible to get confidence score(probablities) at    frame  level  in  htk.  If so  could  you  please explain me the same.

My dictnory structure:  I dont have word level dictornay.. my dictnory is just phones  i.e. it looks like this :

a     a

aa   aa

b     b

and so on ....

Thanks in advance. Your help highly appreciated.

 

--- (Edited on 7/23/2013 10:06 am [GMT-0500] by manju) ---

PreviousNext