This paper presents a framework for VOIP database generation and an investigation of the impact of VOIP characteristics on the accuracy of automatic speaker identification system. Exactly we study the impact of G711 and iLBC codec, and the influence of packet loss. A set of experiments are done on the generated databases to find the best feature extraction method for speaker identification on VOIP. The acoustic features considered are the most commonly used ones: MFCCs, LPCs and PLPs. Speaker models used in this study are based on Gaussian Mixture models and are implemented using HTK. VOIP databases used for training and testing are created using Asterisk.