Plínio Almeida Barbosa, University of Campinas
This work presents an automatic prosodic salience detector algorithm which does not require the use of language-specific duration values. It is implemented in two steps: automatic detection of vowel onsets (VO) followed by the detection of normalized VO-to-VO duration peaks. The algorithm's performance is compared to that of a semi-automatic version. Perceived salience is also compared. For both fast and slower read speech, precision and accuracy of perceived word salience are between 61 and 80\%. In a larger corpus of read and storytelling speech, precision is generally higher than 70\%, whereas accuracy is higher than 80\% when the automatic version is compared with the semi-automatic one. The automatic algorithm's performance is found to be similar to that of the prominence detector reported in Wang and Narayanan (2007).