Interpolation of Estimators Derived from Sparse Data
Original Publication Date: 1981-Sep-01
Included in the Prior Art Database: 2005-Feb-12
An estimation technique is described for building statistical speech models wherein improved estimators are calculated from sparse data in the speech recognition application. Markov sources are a widely used class of statistical models of generators of data. A Markov source consists of a set of states S, transitions T between them, and outputs from alphabet A associated with the transitions. At given time instants, the source changes state from s to s' along some transition t (with probability q(s)(t)), and outputs the letter a associated with t. The usual modeling approach is to determine (by intuition, physical reasoning, etc.) the structure S, T, A of the source and then use sample training data to estimate the transition probabilities q(s)(t). An efficient iterative method of estimation due to Baum  exists.