Unsupervised learning of HMM topology for text-dependent speaker verification