Bimodal Speech Recognition Based on Hierarchical Parallel Boosting

In this paper, a weak learning algorithm is boosted to a strong effective learning algorithm using the boosting algorithm. The traditional boosting algorithm costs a great deal of running time. In order to decrease the running time, a hierarchical parallel boosting algorithm is proposed, which supports the multi-class classifying problem. Some simple classifiers using this algorithm can be trained in parallel. In addition, within one classifier, models for every class can be trained in parallel, too. Experimental results of bimodal speech recognition show that the new algorithm is able to produce classifiers as accurate as the traditional boosting classifier with the same number of base classifiers, but with greatly reduced running time.