My first neural network
This is my first neural network. I wanted it to be as simple as possible but still wanted to do something. So I went for the most simple feedforward neural network. Wikipedia does not make this clear but all it is is two lines...
The bold font signifies these matrices or vectors. The superscript in brackets is to signify both the Ws and bs are different. Lower case bold fonts are vectors while uppercase bold fonts are matrices. The f is called the sigmoid function and makes sure the outputs are between zero and one.
For this application I used Armadillo Which takes care of all the vectors and matrices. That means in this C++ application it is still just two lines...
h_vec=sigmoid(W1*x_vec+b1_vec);
v_vec=sigmoid(W2*h_vec+b2_vec);What it does
It recognizes two words spoken by me, "on" and "off".
Click the Video to see it in action.
Training
Training is the difficult thing with neural networks. I used Matlab to train it to my voice using snippets of audio of me saying "on" and "off" That I cut and pasted together. I then exported the trained matrices so that the C++ program could import the matrices.
So using Matlab or Octave you should be able to do the same thing with your voice.
Finally
Hopefully I'll write a little more about this later. Neural networks are fascinating as they seem almost alive. Watching my network learn was so awesome. I felt like a proud parent.
Jonti