Speech recognition based on variable information rate model

The authors propose a variable information rate model to show the importance within each portion of a speech signal. The information rate parameters used are discriminative frame weights which depend on both neighbouring feature vectors and on the corresponding model. The information rate parameters are found using the generalised probabilistic descent method. In speaker-independent speech recognition experiments, the proposed method results in a considerably improved performance, compared to the conventional method which treats all speech segments with the same importance