Generating complementary systems for speech recognition

Large Vocabulary Continuous Speech Recognition (LVCSR) systems often use a multi-pass recognition framework where the final output is obtained from a combination of multiple models. Previous systems within this framework have normally built a number of independently trained models, before performing multiple experiments to determine the optimal combination. For two models to give improvements upon combination, it is clear that they must be complementary, i.e. they must make different errors. While independently trained models often do give improvements when they are combined, it is not guaranteed that they will be complementary. This paper presents a new algorithm, Minimum Bayes Risk Leveraging (MBRL), for explicitly generating systems that are complementary to each other. This algorithm is based on Minimum Bayes Risk training, but within a boosting-like iterative framework. Experimental results are reported on a Broadcast News Mandarin task. These experiments show small but consistent gains when combining complementary systems using confusion network combination.