Blind estimation of hidden Markov models

In this paper, an on-line blind parameter estimation scheme for hidden Markov models (HMMs) is developed. The parameters to be estimated in the paper include state transition probabilities, observation vector and measurement noise density. Some implementation aspects of the proposed blind estimation algorithm are discussed. Computer simulations show that our algorithm can converge to the true values under different noise environments and initialisations. Furthermore, it can track the slowly varying changes of HMM's parameters.