Detection of syntactic boundaries by partial analysis-by-synthesis of fundamental frequency contours

A phrase boundary detection experiment was realised using a previously proposed method for the use of prosodic features in automatic speech recognition. The method consists of comparing the observed fundamental frequency contour of an utterance with the contour generated by a model for each recognition hypothesis, using a scheme of partial analysis-by-synthesis. Allowing l-mora discrepancies, the detection rate reached 95%. To cope with the high variability in the errors obtained by the partial analysis-by-synthesis even for correct hypotheses, further experiments were carried out in two ways: smoothing the observed contour with a low-pass filter to reduce the effects of microprosody, and adapting the initial conditions of the model-generated contours to the speaker. Experiments were also conducted applying the method at non-phrasal boundaries with the objective of obtaining clues that permit the differentiation between phrase boundaries and non-phrasal boundaries