Dealing with Out-of vocabulary Words and Filled Pauses in Word N-gram Based Speech Recognition System

摘要

For practical use of spoken dialog systems and dictation systems, it is important to cope with out-of vocabulary words and filled pauses including the phenomena such as interjection, restart and hesitation. To address these problems, this study tries to use an unknown-word processing (UWP) method for a word N-gram language model based continuous speech recognition system. We investigate an UWP method which employs a subword sequence decoder with subword acoustic models to produce unknown-word hypotheses. This method has been shown to be effective on a small vocabulary task tested with a context-free grammar-based recognition system. This paper proposes an efficient method for incorporating the UWP into a word N-gram language model--based recognition system. We performed a series of experi- ments to show the effectiveness of the method for spoken dialog tasks and a dictation task. The experimental results show that a semantic accuracy was improved by 48/100 using the UWP method. Also, in compared with the result of a system using context-free grammar, the word N-gram based system could further improve the semantic accuracy for spontaneous speech. Furthermore, we performed a recognition experiment for a large-vocabulary dictation task. As a result, although only a slight improvement was observed in terms of the word accuracy, the high performance for detecting the existence of unknown-word in an utterance could be achieved.