word boundary- hypothesisation in hindi speech - Speech and ...

were asked to identify the boundaries between the words in the phrase. The results
showed that the listeners were able to detect many word boundaries better than
chance. An analysis of the resillts showed that stress and rhythm were the primary cues
used by the listeners to detect worcl boundaries.
Similar studies on the use ad prosodic cues were performed for other languages
also. In a study on Japanese speech [Nakagawa and Sakai 19791, utterances of one to
three word sequences of city naimes and digits was synthesised. When the voiced
segments in the synthesised speech were replaced by sinusoids and the unvoiced
segments by white noise with both pitch and energy unaltered, 95% of the word
boundaries in the sequences were still recognised by the listeners. When either energy
or pitch is modified, the recognitioln rate dropped to 92%.
In a later study [Lea 19801, an algorithm was developed to detect major
syntactic boundaries in English speech using falllrise patterns in the pitch contour. The
algorithm looks for substantial decreases in FO (7% or more) followed by an increase
(7% or more). It then marks a boundary at the last of the lowest FO values in the
valley. Application of this algorithm on a data of 230 sentences, resulted in locating
nearly 90% of the major syntactic boundaries correctly with false alarm rate between 5
to 10%. In the same study it was suggested that some word boundaries can also be
detected by using durational changes caused by 'prepausal lengthening'. It was also
suggested that by detecting pauses in speech, which occur as long silences, one can
detect a few word boundaries.
The relation between the prosody and the syntactic structure of a sentence was
also examined in some recent studies. It was shown that boundary tones, which are
distinctive changes in pitch, signal major boundaries in phrases and sentences
[Beckman and Pierrehumbert 19861. In a recent study, it was found that abrupt changes