Understanding Spontaneous Speech

UNDERSTANDING SPONTANEOUS SPEECH Wayne Ward 1 Carnegie Mellon University Computer Science Department Pittsburgh, PA 15213 ABSTRACT When speech understanding systems are used in real applications, they will have to deal with phenomena peculiar to spontaneous speech. People use language differently when they speak than when they write. Spoken language contains many interjections, filled pauses, etc. Speakers often don't use well-formed sentences. They speak in phrases, have restarts, etc. Systems designed for written or read text will encounter serious difficulties processing such input. This paper outlines our strategy for dealing with spontaneous spoken input in a speech recognition system. INTRODUCTION As systems become more habitable and allow users to speak naturally, speech recognizers and parsers are going to have to deal with events not present in written text or read speech. Spontaneous speech contains a number of phenomena that cause problems for current systems. • filled pauses - noises made by the speaker that don't correspond to words (ah, uh, um, etc). • restarts - repeating a word or phrase. The original word or phrase may be complete or truncated. • interjections - extraneous phrases as in "on line thirty, I guess it is". • unknown or mispronounced words • ellipsis • ungrammatical constructions - Users make errors of agreement (sub-verb, number, etc) and may use constituents in unusual orders ("to the utilities cell add fifty dollars"). These phenomena violate constraints currently used by speech recognizers to increase performance. This can cause complete recognition failure for an utterance. In his paper on habitability, Watt (1968) characterizes the problem as a difference between COMPETENCE and PERFORMANCE. We must recognize what people say, not what they think is grammatical. In real dialogs, much can be understood from context and is left out of utterances. Ellipsis is very common. Many elliptical utterances are not just deletions from expected well-formed sentences. Consider the utterances "okay .. expenses .. mortgage seven forty eight point fifty seven .. car payment, two forty three, point twenty seven, bank surcharge, fifteen dollars". The focus is the information to be transferred, a label specification and an amount. Each utterance is the simplest expression of the neccessary information with no other embroidery. The solution to this problem must involve both parsing and recognition strategies. It must resolve the competeing aims of reducing search space and remaining flexible to the unexpected. Our approach is a combination of specific modelling of acoustic properties and a flexible control structure. 1This research was sponsored by the Defense Advanced Research Projects Agency (DOD), ARPA Order No. 5167, under contraact number N00039-85-C-0163. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency or the US Government. 137 LIMITATIONS OF CURRENT RECOGNIZERS Current state-of-the-art speech recognition systems make several assumptions about the input in order to increase performance: • A single well-formed sentence is spoken. Well-formed means acceptable to the system's language model. • …

Content truncated to 500 words