Correlatos fonético-acústicos de fronteiras prosódicas na fala espontânea

Abstract

The present study presents proposals for automatic prosodic boundary detection models, based on multiple phonetic-acoustic parameters, in order to enable the construction of a computational tool capable of segmenting spontaneous speech into intonation units. A seven-segment sample of spontaneous monological speech was segmented into intonation units by 14 trained annotators. The prosodic boundaries perceived by the annotators were marked as terminal and non-terminal prosodic boundaries. A Praat acoustic analysis software script was developed to extract a series of phonetic-acoustic parameters along the sound signal. Two statistical classifiers, Random Forest (RF) and Linear Discriminant Analysis (LDA), were applied to generate models comprising subsets of phonetic-acoustic parameters, which could function as prosodic boundary predictors. An initial evaluation of the classifiers indicated that both display relative success in automatic boundary detection, although the LDA classifier presented a higher percentage of accuracy in predicting prosodic boundaries. Therefore, the models obtained through the LDA were refined. The final model for the automatic detection of terminal boundaries presents an 80% convergence in relation to the boundaries identified by the annotators for the speech sample. Concerning the non-terminal boundaries, three classification models were obtained. The sum of the number of boundaries identified by the three models corresponds to a 98% convergence in relation to the boundaries marked by the annotators.