A corpus-based learning method for prominence detection in spontaneous speech

Abstract

The aim of this paper is to present a software tool called ANALOR, which allows semi-automatic prominence detection in spontaneous French. On the basis of a manual annotation performed by two experts on a 70-minute long corpus including different regional varieties of French (Belgian, Swiss and metropolitan French) and various discourse genres (from read speech to spontaneous conversations), our system conducts a learning-method in order to determine the best thresholds for prominence prediction. This procedure appreciably improves detection, with consistency between automatic identification and the human labeling rising from 75.3 without training to 79.1 of f-measure after corpus-based learning.