A minimum cross-entropy approach to hidden Markov model adaptation

摘要

An adaptation algorithm using the theoretically optimal maximum anposteriori (MAP) formulation, and at the same time accounting fornparameter correlation between different classes is desirable, especiallynwhen using sparse adaptation data. However, a direct implementation ofnsuch an approach may be prohibitive in many practical situations. Wenpresent an algorithm that approximates the above mentioned correlatednMAP algorithm by iteratively maximizing the set of posterior marginals.nWith some simplifying assumptions, expressions for these marginals arenthen derived, using the principle of minimum cross-entropy. Thenresulting algorithm is simple, and includes conventional MAP estimationnas a special case. The utility of the proposed method is tested innadaptation experiments for an alphabet recognition task