This study investigates the relationship between sentence prominence and the predictability of word-specific statistical descriptors of prosody. We extend from an earlier word-invariant model by studying a model that marks words as prominent if the acoustic prosodic features differ from their expected values during the lexemes. To test the approach, the most common acoustic features associated with the perception of prominence are extracted and several lexeme-specific statistical measures are computed for each feature. Simulations are conducted on a corpus of continuous English speech and the algorithm output is compared to manually assigned prominence labels. The results show that the deviant prosodic descriptors of the words correlate with the perception of prominence. However, this effect is much smaller than that obtained by modeling the prosodic predictability at the utterance level, suggesting that context-independent lexeme-specific models are unable to capture relevant aspects of sentence prominence.