Improved efficiency of pruning accelerates the search process and leads to a more time efficient speech recognition system. The goal of this work was to develop a new pruning technique which optimizes the well known probability-based pruning (beam width) by utilization of confidence measurement. We use normalized hypotheses scores to guide the beam width of the pruning process dynamically frame per frame during the whole utterance. Compared with classical pruning techniques like fixed beam pruning and histogram rank pruning we achieved significantly better results concerning the time consumption of the recognizer. The speed of the recognition process could be accelerated up to 14 times with a slight degradation in recognition accuracy.