Speech intelligibility (SI) plays an important role in second language learning. It can be influenced by many factors and various approaches have been explored to measure it. In this study, the intelligibility of Dutch non-native speech was measured by Visual Analogue Scale (VAS) and word accuracy (W_Acc) that was automatically derived from orthographic transcriptions (OTs). A large number of acoustic-phonetic features were automatically extracted from the audio files. To deal with the multicollinearity issue, feature reduction through different regression approaches was investigated. The results revealed that, as a whole, the LASSO regression approach outperformed (highest R^2 at 0.52, lowest RMSE and MAE) the other explored regression methods. The regression models predict the intelligibility measures assigned by human raters well. The obtained findings indicate the usability of automatically extracted acoustic-phonetic features to provide a basis for the assessment of SI.