Structural learning of dynamic Bayesian networks in speech recognition

We present a speech modeling methodology where no a priori assumption is made on the dependencies between the observed and the hidden speech processes. Rather, dependencies are learned form data. This methodology guarantees improvement in modeling fidelity compared to HMMs. In addition, it gives the user a control on the trade-off between modeling accuracy and model complexity. Furthermore, the approach is technically very attractive because all the computational effort is made in the training phase.