Predicting speech fluency in children using automatic acoustic features

The present study aims at predicting the speech fluency of children using automatic acoustic measures derived from forward-backward divergence segmentation (FBDS). Thirteen Korean children were recorded while reading out loud a set of sentences. Three native-Korean speakers evaluated the fluency of each sentence on a five-point scale. A FBDS algorithm was used to segment speech recordings into sub-phonemic units and silent segments. In addition to the low-level acoustic features directly derived from FBDS segments, higher-level acoustic features were computed by clustering FBDS segments into pseudo-syllables and silent breaks. Both low- and higher-level features were used to predict average ratings of speech fluency, using a leave-one-speaker-out cross-validation scheme and three regression models: a multiple linear regression, a support vector regression, and a random-forest regressor. Highly accurate predictions were achieved, with average root-mean-square errors (RMSEs) as low as 0.3. Prediction accuracy did not significantly change as a function of regression model. Using higher-level features yielded lower RMSEs than using raw FBDS features. The results of a multiple linear regression using higher-level features (R2 = 0.94) suggest that speech/silence ratio and pseudo-syllable rate are the two most important predictors of speech fluency.