Automatic Detection of Prosodic Boundaries in Brazilian Portuguese Spontaneous Speech

This paper presents some models based on multiple phonetic-acoustic parameters for the automatic detection of prosodic boundaries in spontaneous speech. A sample with seven excerpts of monologic Brazilian Portuguese spontaneous speech was segmented into prosodic units by 14 trained annotators. The perceived prosodic boundaries were annotated as terminal or non-terminal prosodic boundaries. A Praat script was prepared in order to extract a set of acoustic parameters during the speech signal. Two statistical classifiers, namely Random Forest e Linear Discriminant Analysis, were used to generate models of subgroups of acoustic parameters that could work as predictors of prosodic boundaries in comparison with the human annotators. The initial evaluation of the classifiers showed that both present relative success in detecting boundaries. The LDA performed better in predicting boundaries and therefore its models were refined. The final model for terminal boundaries showed 80% of agreement with human annotators. As for non-terminal boundaries, three models were obtained. The sum of boundaries identified by the three models together corresponds to an agreement of 98% with the human annotators.