Prosodic scoring of word hypotheses graphs

Abstract

Prosodic boundary detection is important to disambiguate parsing, especially in spontaneous speech, where elliptic sentences occur frequently. Word graphs are an efficient interface between word recognition and parser. Prosodic classification of word chains has been published earlier. The adjustments necessary for applying these classification techniques to word graphs are discussed in this paper. When classifying a word hypothesis a set of context words has to be determined appropriately. A method has been developed to use stochastic language models for prosodic classification. This as well has been adopted for the use on word graphs. We also improved the set of acoustic-prosodic features with which the recognition errors were reduced by about 60% on the read speech we were working on previously, now achieving 10% error rate for 3 boundary classes and 3% for 2 accent classes. Moving to spontaneous speech the recognition error increases significantly (e.g. 16% for a 2-class boundary task). We show that even on word graphs the combination of language models which model a larger context with acoustic-prosodic classifiers reduces the recognition error by up to 50 %.