Dynamic beam pruning strategy using adaptive control

In large vocabulary continuous speech recognition, huge search space results in vast computational cost. While most pruning search strategies can reduce the computation, but the recognition rate often decreases. This paper aims to reduce the computation time without any sacrifice of the recognition rate. By means of the adaptive control theory, a novel pruning method is presented. It can automatically steer beam to make search space attain an expected size. In order to make the decoder more efficient, the average number of active model instances is used as the dynamic reference of the adaptive system. Compared with the baseline system with fixed beam and histogram pruning, the proposed method leads to a significant reduction in the computation time and a slight improvement in word accuracy.