Classification of real-world data poses a number of challenging problems. Mismatch between classifier models and true data distributions on one hand and the use of approximate inference methods on the other hand all contribute to inaccurate classification. Recent work on boosting by Schapire et al. and additive probabilistic models by Hastie et al. have shown that improved classification can be achieved by linearly combining a number of simple classifiers. Building upon this spirit, we present a Bayesian network-based framework for mixing multiple classifiers. We also analyze the bound on the generalization error for this combined classifier. We give results on some standard datasets and demonstrate its usefulness in a real-world task of multimodal speaker detection where we improve upon performance of a more complex Bayesian network model. Improved results indicate the significant potential of the Bayesian network of classifiers approach.