A hidden Markov model with controlled non-parametric emissions

Abstract

A novel nonparametric model is introduced to model and control emission densities of a non-ergodic hidden Markov model. Having both multiclass and one-class classifications simultaneously, for recognizing the best match between multiple classes and then accepting or rejecting the given input pattern, is the major characteristic of this algorithm. Also, since the proposed method creates independent feature spaces and trains by positive samples only, it allows the vocabulary of trained patterns to grow without any concern about growing into a negative set (which is a problem with algorithms that use negative/garbage sets for binary training).