Training data for neural net classifying songs

I am trying to create a neural net that is able to take in feature information of a song such as pitch intervals and then tell me which song it belongs to in the database. The database consists of MIDI files and the input is WAV files.

The input WAV file is processed for onset and pitch detection. Currently I am only using interval (difference between successive pitches) for training and classifying.

My current approach is to make a neural net for each song in the database and train to only know that specific song using a curve fitting NN. For example net(song) = 1 and net(not_song) = 0.

The problem is that the neural net performs badly. I am using the MATLAB neural net tool to train. However, I think that by providing random snippets of the actual pitch intervals obtained form the MIDI file does not make good training data.

I would appreciate any ideas on how I can produce training data that will actually work.