Comparing maximum a posteriori vector quantization and Gaussian mixture models in speaker verification