CNN deep network consist of inbuilt feature extraction (flattening) layer along with classification layers. By omitting the feature extraction layer (conv layer, Relu layer, pooling layer), we can give features such as GLCM, LBP, MFCC, etc directly to CNN just to classify alone. This can be acheived by building the CNN architecture using fully connected layers alone. This is helpful for classifying audio data.

http://cs231n.github.io/convolutional-networks/ visit this page for doubts regarding the architecture. I have used C->R->F->F->F architecture