Word-level F0 modeling in the automated assessment of non-native read speech

This study investigates methods for automatically evaluating the appropriateness of F0 contours in the task of automated assessment of non-native read aloud speech. The F0 contour of a test taker’s spoken response is represented as a fixed-dimension vector with a word-level F0 value corresponding to each word in the prompt text. This vector is then correlated with gold standard vectors extracted from native speaker responses. Three different measures are used to describe the F0 contour within a word, including the mean of the F0 values, the difference between the mean values for each word and its neighboring words, and polynomial regression parameters. Additionally, features are developed based on a human expert’s annotations, in which different types of words in a reading passage are identified as prosodically more important than others. Experimental results demonstrate the effectiveness of applying the proposed features to the automated prediction of intonation and stress scores for non-native read aloud speech.