A parallel hybrid system of HMM and GMM modeling techniques was implemented and used in a telephony speaker verification and identification system. Spectral subtraction and Weighted Projection Measure were used to render this system more robust against additional noise. Cepstral Mean Subtraction method was also applied for the compensation of convolution noise due to transmission channel degradation and differences in the frequency response of telephone handsets. For a population of 100 speakers of FARSDIGITS1 database with a SNR of 8.8 dB, a speaker identification performance of 95.51% and a speaker verification error rate of 0.37% were obtained. Several score normalization methods in utterance and frame level and weighting of model scores were also implemented, and then compared and evaluated. It was shown that these methods improve discrimination between speakers and yield a reduction of speaker verification and identification error rates.