Dependency analysis of read Japanese sentences using pause and F0 information: a speaker independent case

This paper deals with the problem of exploiting prosodic information in syntactic analysis of sentences. Duration of pauses at phrase boundaries and relative F0 contour features have been found to be effective for parsing in speaker-dependent case. In this paper, effectiveness of pause and F0 information was examined in a speakerindependent manner by using prosodic features extracted from the spoken version of sentences of 44 speakers. By simplifying the estimation of both pause and F0 contour models, better performance was obtained. Linear combination of pause and F0 information gave significant improvements in parsing accuracy. It was also shown that the effectiveness of pause information was larger when pause models were estimated separately for zero-duration and non-zero-duration pauses.