Techniques for information retrieval from speech messages
Richard Cameron Rose · 1991
• The goalin speech-message information retrieval is to categorize an input speech utterance according to a predefined notion ofa topic, or message class. The components ofa speech-message information-retrieval system include an acoustic front end that provides an incomplete transcription ofa spoken message, and a message classifier that interprets the incomplete transcription and classifies the message according to message category. The techniques and experiments described in this paper concern the integration ofthese components, and represent the first demonstration ofa complete system that accepts speech messages as input and produces an estimated message class as output. The promising results obtained in information retrieval on conversational speech messages demonstrate the feasibility ofthe technology. THE GOAL IN SPEECH-.MESSAGE information retrieval is similar to that of the more well-known. problem ofinformation retrieval from text documents. Text-based information-retrieval systems sort large collections of documents according to predefined relevance classes. This discipline is a mature area ofresearch with a number of well-known document-retrieval systems already in existence. Speech-message information retrieval is a relatively new area, and work in this area is motivated by the rapid proliferation ofspeech-messaging and speech-storage technology in the home and office. A good example is the widespread application of large speech-mail and speech-messaging systems that can be accessed over telephone lines. The potential length and number of speech messages in these systems make exhaustive user review ofall messages in a speech mailbox difficult. In such a system, speech-message information retrieval could automatically categorize speech messages by context to facilitate user review. Another application would be to classifY incoming customer telephone calls automatically and route them to the appropriate customer service areas [1]. Unlike information-retrieval systems designed for text messages, the speech-message information-retrieval system illustrated in Figure 1 relies on a limited-vocabulary acoustic front end that provides only an incomplete transcription of a spoken message. The second stage of the system is a message classifier that must interpret the incomplete transcription and classifY the message