Large Margin Hidden Markov Models in command recognition and speaker verification problems

Discriminative properties of different HMM structures, parameters and training algorithms are analyzed in the task of isolated words recognition (digits and robot controlling commands) and speaker verification. The ergodic, Bakis and chain HMM structures are considered, having constant or variable number of states. The classical Baum-Welch training algorithm is compared with the discriminative training, using the large margin approach. The class separation is increased by using the proper HMM structure, the variable number of HMM states and a large-margin HMM training algorithm, based on the extension of the training sequence.