Methods for combining the information of various features in speech recognition

In studies of speech recognition based on features, it is often discovered that the recognition performance of some features for some utterances is better than for the others while the other features have opposite effects. They have, to some extent, the complementary relation in the recognition of some utterances. Based on HMM which is now widely used in speech recognition studies, three efficient methods for combining multiple complementary features to improve the recognition performance of HMM are presented in this paper. They are defined as the maximum parameter method, the all parameter method and the most reliable parameter method. The three methods respectively improve the performance in multi-speaker Chinese digit DHMM/VQ recognition from 89% to 92.3%, 95.7%, 94.3%. This paper describes the three methods and gives experimental results and analysis in multi-speaker Chinese digit DHMM/VQ recognition in detail.