Automatic measurement of vowel duration via structured prediction