How to create a dataset of neural network target?

[fikrul]

hello,

im going to use a back-propagation neural network for my voice recognition. My question is how to create a set of dataset consisting of my target? Let say i have 3 target of word 'a', how im going to aplly those 3 target in my newff function?

Please help me on this matter..

 

[jabi]

Hi
u must extract some feature , and then use them as input of NNT.
also u must select some of them as train and test data.
if u have 3 target, your NNt has 3 output that for first class, first output must set as .98 and .02 for two else, and so on.

spouse
X=input_data; matrix size is N*M that M is length of features and N is words Number that created from 3 class,
Y=Output ; matrix size is N*3; => output of first class: [.98; .02;.02], second class: [.02; .98; .02],...

net = newff(X,Y,H_L); H_L is length of Hidden layer (is greater if input length)

net = train(net,X,Y); for train
Y = sim(net,P); for test

 

[fikrul]

helo jabi,

so far, i already had done feature extraction for the input. The size of that input is (479 x 13).

For the target, you said:

"if u have 3 target, your NNt has 3 output that for first class, first output must set as .98 and .02 for two else, and so on."

meaning that, if i have 1 target, my NN should have 1 output right?
1) My question is how do we set the output as u mentioned before for example .98 and .02. Do i need to pick any value randomly?

2) What is output classes?

Please help me,,im really stuck with definining the target for my NN..

 

[jabi]

Hi fikrul

If u want using NNt for classification, U have more than one output (number of your classes), and value of NNt in each output, represent the similarity of input pattern to each of classes
.
notice that we use Transfer functions in output of each of each layer; then if u use log Sigmoid (logsig - in matlab), ur output will be limited to (0-1).
when u test a input vector, each output that have bigger value, is your input class.( N-th output represent the N-th class)

because of using logsig as transfer function, we use .98 and .02 to create our output vectors.

Helpful?

 

[fikrul]

hi jabi,

so far, its helpful and i do understand about the newff function.i just got problem and confusing at the beginning part,

1) for the NNt input, i used the input signal that i've been extracted which is from mfcc, is it correct?

2) for NNt target, do i need to extract the voice signal first before use it as a target for my NNt?

If u dont mind, may i ask your email so that i can easily discuss with you..?
this is my email: (def_cult@yahoo.com)