A generalization of the maximum a posteriori training algorithm for mixture priors

Prior information about the operating environment of a speech recognizer is often general and abstract. Frequently information such as the number of speakers with foreign accents or the number of callers using cellular phones is readily available. Incorporating this information during model training is difficult. This paper generalizes the popular MAP training algorithm derived by Gauvain and Lee (1994) so that more prior information can be utilized during training. The priors are in the form of mixture distributions with each mixture component representing a unique property of the data and the mixing weights defined by the a priori constraints. Using the training algorithms derived here it is shown that significant performance improvements can be obtained.