Speech recognition with neural network pre proccessing problem

I want to do a speech recognition project with neural network. I have used MFCC for feature extraction and I used its coefficient for neural networks input. My inputs are numbers from 0 to 9 and target vector is t=0:10:90 . After train network with two hidden layyer I obtained for correct result for my input matrix. İt is like if input =1 then output is between 10 and 20 and so on. But when I tried to add new data it gives wrong result because of time delay during recording sound. So I thougth ,I could use covariance coefficient for determining time delay. I found time delay. but I recognize I dont have referance signal for new data. Its mean I have 9 different signal but I dont know which one is I need to corrolate. Maybe I can corrolate my new data with all of them , then neural network will loss its necessary. Following graph is two signal which is 6 and they have time delay also. So as you see upper , I confused a little. What is the solution to solve this time delay problem ?

[Arif Izudin]

can i get the full coding?

Answer 1

[Greg Heath]

Insufficient quantitative detail.

It seems to be a classification/pattern-recognition problem with 10 classes/categories. Therefore use patternnet with one hidden layer. The columns of the target matrix should be columns of the 10-dimensional unit matrix eye(10) and

 trueclassindices = vec2ind(target)
 target           = ind2vec(trueclassindices)

See the patternnet documentation

 help patternnet
 doc patternnet

Practice classification on the MATLAB classification/pattern-recognition examples obtained from

 help nndatasets
 doc nndatasets

Also see of my posts using the search words

 greg patternnet

Now all you have to do is figure out how to preprocess the input waveforms so that members of the same category look similar.

Only two examples are shown above. Low-pass filtering, normalization, shifting and truncation (not necessarily in that order) might be useful. One way to deal with shifting is to use the magnitude of the fft.

Hope this helps.

Thank you for formally accepting my answer

Greg

[aliha wasim]

hello! Can you help me in mfcc if you already implement it. .I am using the melcepst code available in voice box but it continuously giving me undefined variable 'w' even though its defined in the code.