Improving ASR processing of ungrammatical utterances through grammatical error modeling

Automatic speech recognition (ASR) of non-native utterances with grammatical errors is problematic. A new method which makes it possible to better recognize such utterances is presented in the current paper. It can be briefly summarized as follows: extract error patterns automatically from a learner corpus, formulate rewrite rules for these syntactic and morphological errors, build finite state grammars (FSGs), and use these FSGs as language models in ASR systems. All rules used in isolation and in different combinations yield lower word error rates (WERs).