Investigations on ensemble based semi-supervised acoustic model training

Abstract

Semi-supervised learning has been recognized as an effective way to improve acoustic model training in cases where sufficient transcribed data are not available. Different from most of existing approaches only using single acoustic model and focusing on how to refine it, this paper investigates the feasibility of using ensemble methods for semi-supervised acoustic modeling training. Two methods are investigated here, one is a generalized Boosting algorithm, a second one is based on data partitions. Both methods demonstrate substantial improvement over baseline. More than 15% relative reduction of word error rate was observed in our experiments using a large real-world meeting recognition dataset.