There has been much interest in new deep learning approaches for representing and extracting high-level features for audio processing. In this paper convolutional deep belief network was used to generate new speech features for text-independent speaker verification. Structure and parameters of a convolutional deep belief network were described. New high-level speech features were extracted using proposed method. Relevance of speaker verification systems for mobile authentication was considered. Gaussian mixture model and universal background model speaker verification system used for experiments was described. Speaker verification accuracy using extracted features was evaluated on a 50 speaker set and a result is presented. Different layers and combinations of layers of convolutional deep belief network were used as a features for a text-independent speaker verification. High level features extracted by convolutional deep belief network were illustrated and analyzed. Reasons of insufficient verification accuracy were described. High-level features extracted by the third layer could be used for gender recognition.