Prosodic boundary detection using syntactic and acoustic information

This paper presents a two-stage procedure for automatic prosodic boundary detection in Russian based on textual and acoustic data. The key idea of the method is (1) to predict all potential prosodic boundaries based on syntax and (2) among these potential boundaries, to choose those which are marked acoustically. For the first stage we developed a system which predicted a potential boundary whenever two adjacent words were not connected with each other in terms of syntax; for this we used a dependency tree parser and added several simple rules. At the second stage we run a random forest classifier to detect the actual prosodic boundaries using a small set of acoustic features. Of all the observed prosodic features pause duration worked best, and for some speakers it could be used as the only acoustic cue with no change in efficiency. For other speakers, however, other features were useful, such as tempo and amplitude resets or F0 range, and the choice of the features was speaker-dependent. In the end the procedure worked with the F1 measure of 0.91, recall of 0.90 and precision of 0.93, which is the best published result for Russian.