The pronunciation of a word can vary widely, and many factors are known to affect this variation. This paper focuses on the role of predictability on word duration. Previous research has suggested that more frequent words are shorter, as are words which are more predictable from neighboring words. This research has tended to focus only on extremely high frequency function words; previous research on content words has not been able to examine natural speech and control for key confounding factors like accent and rate of speech. We examined 1401 content words from the Switchboard corpus and studied the role of word frequency, conditional and joint probabilities with neighboring words and a measure of word semantic association, Latent Semantic Analysis. Using multiple regression to control for confounding factors, we show that two predictability variables principally affect a word’s duration—word frequency, and the conditional probability of a word given the following word. Other predictability variables do not make an additional significant contribution. We discuss the implications for pronunciation modeling.