Reinforcement of MLLR speaker adaptation using optimal linear interpolation

An optimal linear interpolation scheme applied to MLLR speaker adaptation is presented. The proposed reinforced MLLR method, called αopt-MLLR, regulates the influence of MLLR adaptation when the training data from a new speaker is improper by adequately incorporating prior knowledge of the initial models into adaptation, and thus ensures the robustness of MLLR adaptation. The proposed mechanism is conceptually simple and computationally inexpensive, and is superior to FLC-MLLR in regard to lower computing cost.