Towards Finding Optimal Features of Perceived Prominence

In this paper we present acoustical as well as lexical features for classification purposes of perceived prominence in read aloud Dutch sentences. Via a perception experiment with 10 naive listeners we derived prominence labels at the word level for 500 sentences. Part of these sentences are used for lexical/syntactical analyses. It turns out that most of the function words are never perceived as prominent, and that specific content words namely adverbs, nouns and adjectives are almost always perceived with some degree of prominence, whereas verbs form a middle class. So we decided to concentrate on the lexically stressed syllables of content words, because these are the words whose prominence is not uniquely classified by their lexical class. In this paper we use F0 range per syllable, both ‘raw’ and corrected for the declination line, to distinguish between the most prominent and non-prominent content words, although intensity and duration features can be used as additional features to improve the classification. As an initial result we can conclude that F0 range is a very good feature to distinguish between prominent and non-prominent content words.