Hello Everyone,
I'm developing an automatic Urdu Speech Recognizer. I've taken 28 samples(7 males, 7 females ,2 recordings from each). After applying filters (Spectral subtraction, Silence Removal), DTW algorithm i calculated 42 MFCC features and then applied K mean algorithm on it.
Now i input the obtained vector to neural network as below :
e=0.2; % initialize error value
net=newff(minmax(input),Tar,[2 5],{'tansig','logsig'},'traingdx');% 2 is number of hidden layer neurons and 5 is number of output layer neuron as i have to classify 5 words. Tar is Target
net.divideParam.trainRatio = 0.7; % ANN will take 70% data for training and 30% for testing
net.divideParam.testRatio = 0.3; net.trainParam.epochs = 1160; % Maximum epochs
net.trainParam.goal =mean(var(Tar')')/100;while e>= 0.0260[net,T] = train(net,input,Tar);test = sim(net,input)temp=round(test)e=Tar-test;disp('error');e=mse(e)end
Now The confusion matrix give me 99.3% result, but when i test the data on the same samples it give me wrong answers.What should i do to get write results and why it is giving incorrect results even after 99.3% of training?
Best Answer