Integrating Hypotheses of Multiple Recognizers for Improving Mandarin LVCSR Performance

In this paper, we investigate how to improve Mandarin LVCSR performance by integrating multiple hypotheses from recognizers running in parallel. Different recognizers are trained by employing: (1) different phone sets, (2) different front-ends, and (3) different training sets. Nbest hypotheses are merged into a character transition network (CTN) and ROVER is used to select the final recognition decision. Both read and spontaneous speech are tested in the ROVER framework. In comparing with the best individual recognizer in the parallel group, the fully integrated ROVER system achieves a relative Chinese character error reduction of 10.1%.