Text-independent speaker verification with dynamic trajectory model

A novel approach is presented for text-independent speaker verification. Based on the frequencies of occurrence of Gaussian component strings in the quantized acoustic trajectories, a universal background trajectory model and multiple target trajectory models are created for the background and target speakers separately. Analysis of the speaker entropy in the trajectory space demonstrates that the segmental dynamic trajectory catches speaker-specific information. Experiments are conducted on the telephony speech used in the NIST 1999 speaker verification evaluation and show that the bicomponent strings achieve good performance and may provide complementary information to a general speaker verification system.