Distance measures for text-independent speaker recognition based on MAR model

For text-independent speaker identification and verification based on the multivariate auto-regression (MAR) model, the authors consider two distance measures: the log likelihood ratio (LLR) and the symmetrized likelihood ratio (SLR) measure, which is a symmetric version of the likelihood ratio measure. The results of experiments indicate that the LLR gives better performance than the SLR for longer training data of 5 or more sentences, and the SLR measure is better for shorter training data. When 10 sentences are used for training, identification and verification rates (after likelihood normalization) are almost the same as those obtained by an HMM-based method. The optimum order of the MAR model is 2 or 3, and the optimum frame period is 16 ms