Robust speech recognition directed by extended template matching in dialogue system

In recent years, the research on dialogue systems becomes more and more important with the ever-increasing demands. However, the automatic speech recognizer of dialogue system is not satisfying because of its bad performance for spontaneous/casual utterances. This paper presents a novel extended template matching (ETM) strategy, which imports the filler models of a keyword spotting (KWS) strategy into the template matching (TM) strategy. Because this recognition strategy not only makes use of the context information and the background knowledge by grammar template, but also adopts filler models to match extraneous speech and non-speech signals, it achieves high recognition accuracy and good robustness. The experiments show that the ETM outperforms the TM and the KWS in both the reading style and the spontaneous style testing sets.