Determining sentence pronunciation difficulty for non-native speakers

This paper investigates the features that determine the sentence pronunciation difficulty for Korean speakers of English. We selected three types of features: length, word frequency, and phonemes that Korean speakers generally replace with other phonemes. We used support vector machines and a multiple linear regression model to determine the pronunciation difficulty of given sentences, and measured the results with a five-fold cross validation. We demonstrated that these features could determine sentence pronunciation difficulty with an accuracy and a correlation coefficient sufficient for computer-assisted pronunciation training (CAPT) systems. The combination of all three feature types had the highest accuracy and correlation coefficient in determining sentence pronunciation difficulty. For single features, the length-based feature type was the most accurate in determining sentence pronunciation difficulty. The phoneme-specific feature type also had high accuracy. Length, phoneme, and word features can be used to guide the automatic choice of sentences for CAPT systems that depend on users’ proficiency levels.