Pause As A Phrase Demarcator For Speech And Language Processing

In spontaneous speech understanding a sophisticated integration of speech recognition and language processing is espceially crucial. However, the two modnles are traditionally designed independently, with independent linguistie rules. In Japanese spc.ech recognition the bunsctsu phrase is the basic processing unit and in language processing the sentence is the basic unit. This difference has made it impracticM to use a unique set of linguistic rules for both types of processing. Further, spontaneous speech contains unexpected utterances other than wellformed sentences, while lingnistic rules for both speech and language processing expect well-formed sentences. They therefore fail to process everyday spoken language. To bridge the gap between speech and language processing, we propose that pauses be treated as phrase demarcators and that the interpausal phrase be the basic common p r o c e s s i n g unit. And to treat the linguistic l)henoI~lena of spoken language properly, we survey relevant features in spontaneous speech data. We then examine the effect of integrating pausal and spontaneous speech phenomena into synt~tctic rules for speech recognition, using 118 sentences. Our experiments show that incorporating pansal phenomena as purely syntactic constraints degrades recognition accuracy considerably, while the additional degradation is minor if some filrther spontaneous speech features are also incorporated. 1 I N T R O D U C T I O N A spontaneous speech understanding system accepts naturally spoken input and understands its meaning. hi such a system, speechprocessing and language processiug must be integrated in a sophisticated manner. Itowew:r, the integration is not straightforward, as the two are stndied independently art(/ have different processing units. Moreover, spontaneous speech contains unexpected phenomena, such as hesitations, corrections and fragmentary expressions, which thus far have not been treated in linguistic rules. The most significant concern in speech processing is raising the recognition accuracy. For that purpose, applying linguistic information, e.g. using stochastic models[l l, syntactic rules[2], sen,antic intbrmation[3] and discourse plan@l], is most promising. In a recent Japanese speech translation system[5] b*lnselsu-based syntactic constraints are successfully applied in the speech processing module[6] 1, However, rules reprel A bunsetsu rouglfly cor responds to a phrase and is the next largest unit af ter the word. T h e nunfl)er of words in a phrase ranges f rom I to 14, art(] the m e a n numl)er is al)ont 317]. senting the same constraints cannot be used directly in sentence-based language processing, where the primary concern is to understand sentence meaning. In speech recognition, a sequence of words forms a bunselsu and a set of bunseisus then forms a sentence. In language processing, on the other hand, where the sentence is the basic processing unit, treating the main verh aud its complements is usually the core of processing. For the sentence kaigi ni moshikomi tai no desu ga, meauing 'I would like to apply for the conference,' the processing discrepancy is sketched in Figure 1: