Neural Network can't recognize some utterances

Neural Network can't recognize some utterances

Post by E. Ess » Fri, 24 Aug 2007 05:22:42


Hello,

I'm working on speech recogntion using some kind of neural networks.

I'm trying LVQ function in matlab Newlvq for 300 utterance (10 words,
10 speaker,3 repetition) for training and the same for test.

The problem is in testing the function sim may produce empty matrix
for one or more utterance , they can't recognize the utterance at
all , as we know neural try to find the nearest class for the
utterance , but in many cases it causes empty matrix.

What's the problem ? How can overcome this?

Thanks in advance
 
 
 

Neural Network can't recognize some utterances

Post by zafa » Sat, 25 Aug 2007 04:12:39


Seems like that you are not training the network enough. Also, your
training data seems to be low but if you train the network (reduce
error function), you should not get empty matrix.



Zafar
============================
http://www.yqcomputer.com/
============================

 
 
 

Neural Network can't recognize some utterances

Post by Greg Heat » Sat, 25 Aug 2007 18:45:59


How are you representing words?

What is the dimensionality of the input space?

Hope this helps.

Greg
 
 
 

Neural Network can't recognize some utterances

Post by E. Ess » Sun, 26 Aug 2007 04:49:58

the problem appears in some cases , I means if I change hidden neurons
or learning rate or epochs , it may work ok or produce empty matrix
for one or more utterances.

I'm using MFCC with 24 coefficients.

I think the training data enough 300 utterances for training , but the
problem maybe in LVQ it self , it can't recognize or training well
with these huge data, because when I'm reduce training data to 150
(only 5 words,10 speaker , 3 repetition) for example it work properly
in all cases without problems !!!
 
 
 

Neural Network can't recognize some utterances

Post by Greg Heat » Sun, 26 Aug 2007 05:19:00


If your inputs are being assigned to the closest
cluster, there is no way to get a nonclassification.

Else you are doing something that I don't understand.

I assume the training algorithm has created more
than 1 cluster per speaker. if not, that could cause
unsatisfactory results.

Have you tried using NEWFF or NEWRB?

Hope this helps.

Greg
 
 
 

Neural Network can't recognize some utterances

Post by E. Ess » Tue, 28 Aug 2007 02:05:42


I used the LVQ method as the book says, so how can you explain that it
worked when I reduced training data in all cases and didn't work when
it was large ?!!!!!