``Prosodic Boundary Detection,''
in Prosody: Theory and Experiment. Studies Presented to Gosta Bruce, M. Horne (Ed.), Kluwer, forthcoming.
The perceived grouping of words in speech, i.e.\ prosodic constituent
structure, contributes meaning to the sequence of words that comprise
an utterance by highlighting its information structure. Since prosodic
boundaries are known to be useful in human speech communication, it
is natural to expect phrasing to be useful in computer speech understanding.
This chapter addresses the problem of recognizing symbolic phrase
boundaries given a speech signal, surveying a variety of proposed
algorithms and their applications. In general, a phrase boundary
model represents the mapping from text (generally syntactic structure)
to the acoustic correlates that mark prosodic phrase structure
(pauses, boundary tones, duration lengthening, etc.). However, the
specifics of the model depend on the linguistic theory of phrasing,
the mathematical framework of the model and the application of phrase
boundary detection, as shown in a discussion of options
for phrase detection. Three applications of prosodic phrase
recognition are described here -- prosodic labeling, constraining
speech/language processing search procedures, and prosodic scoring of
word hypotheses -- together with their implications for algorithmic
choices in phrase detection. The chapter concludes with a discussion
of open problems in prosodic phrase recognition.
Return to SSLI Lab Publications