How to extract interesting information for identity verification process from spectrograms?

Nowadays, identity verification support systems are becoming more and more popular. Machine learning is one of the leading fields of research from all over the world. However, each classifier needs a large number of samples to be properly trained. Preparing such samples proves to be a big problem for several reasons. One of them is the quality of the recording, another is the problem of feature extraction. In this work, the idea of processing sound samples by using their graphical representation in the form of spectrograms is described. The process removes specific, redundant information from the samples and then performs feature extraction. The proposed technique has been tested for identity verification using convolutional neural networks. The performed tests and obtained results have been described and discussed to indicate numerous advantages and disadvantages of the proposed technique.