Large Margin Discriminative Semi-Markov Model for Phonetic Recognition

摘要: This paper considers a large margin discriminative semi-Markov model (LMSMM) for phonetic recognition. The hidden Markov (HMM) framework that is often used recognition assumes only local statistical dependencies between adjacent observations, and it to predict label each observation without explicit phone segmentation. On the other hand, (SMM) allows simultaneous segmentation labeling of sequential data based on segment-based Markovian structure among all observations within segment. For which inherently joint problem, SMM has potential perform better than HMM at expense slight increase in computational complexity. considered this non-probabilistic discriminant function linear feature map attempts capture long-range observations. parameters are estimated by learning structured prediction. parameter estimation problem hand leads an optimization with many constraints, constrained solved using stochastic gradient descent algorithm. proposed LMSMM outperformed TIMIT task.