Revising parameters for predicting L2 speech fluency and proficiency

Abstract

The aim of the study was to investigate whether integrating parameters based on pause location improve the prediction of fluency and proficiency in L2 Finnish monologic speech. To answer this question, multiple linear regression models were fitted using two data sets containing L2 Finnish speech and expert assessments of fluency and oral proficiency. Separate models were derived for fluency and proficiency using combined data as well as the two separate data sets. The comparison of the models indicate that pause-by-location parameters can improve the prediction of L2 fluency and proficiency, but the relevant parameters and their significance in the regression models depend on the speech data. Parameters with low incidence work only in longer speech samples, while parameters with frequent occurrence can be used even in shorter samples. The results have implications for improving automatic assessment of L2 speech especially in low-resource languages.