Efficient language model development for spoken dialogue recognition and its evaluation on operator's speech at call centers

摘要

While a language model for recognition of spoken dialogue is ideally built from a very large, specific-task-oriented corpus, a great amount of time and effort is required to develop such a corpus, and this involves both the audio recording and written transcription of large amounts of speech data. Training data for a language model should match the target task in both topic and style. What is needed, then, is a method to utilize previously existing spoken dialogue corpora that are not necessarily related to the specific target-task. Such corpora would be combined with documents related to the topic of the target-task to develop a language model for the target spoken-dialogue. In this paper, we propose a method for combining previously existing corpora with key phrases (i.e. phrases that contain keywords) extracted from task related documents. Even though the added data is from documents related to the target dialogue, since it consists of key phrases, stylistic differences (between document data and the actual dialogue to which the model will be applied) are not a problem. We have produced a model using this method and have evaluated it in use on actual spoken dialogue collected at call centers. Experimental results show that a relative 13% reduction in word error rate could be achieved with the addition of key phrases. This performance is nearly as good as that which would be achieved on the basis of a large, expensive transcript-corpus, and the cost of producing the key phrase data is essentially negligible. Such cost reduction achieved by our method will enable speech recognition applications to be more widely used.