Study on Framework for Chinese Pronunciation Variation Modeling

The pronunciation variations, which badly influenced the performance of ASR system, are serious in continuous speech, especially in spontaneous speech. Many research works are focused on pronunciation variation modeling in recent years. A framework for Chinese pronunciation variation modeling is described in this paper. The main idea is that the pronunciation variations are hidden in the recognition errors and it could be found out and modeled. The linguistic knowledge can be used for supervising the finding variations, and it can be used to model the variations directly. Under this framework, the pronunciation variation modeling can be made more reliably and more expediently than conventional methods. Our primary experimental results show it is effective for Chinese continuous speech recognition.