Real-time speaker identification using speaker model distance

In real-time speaker identification systems, sufficiently fast methods are required. As the number of registered speakers increases, the traditional methods become intractable due to high computational costs. In this paper, a two-step approach is proposed which prunes the speakers´ search space in the first step using the distances between models and then finds the target speaker among the candidate speakers in the second step. This method proved empirically effective and increased the speed of the common GMM method more than 28fold without any decrease in accuracy.