Gaussian Mixture Modeling of vowel durations for automated assessment of non-native speech

This paper investigates using Gaussian Mixture Model (GMM) based vowel duration features for automated assessment of non-native speech. Two different types of models were compared: a single GMM trained on a reference corpus of native speech and separate GMMs for different proficiency levels trained on a large corpus of scored non-native speech. 13 vowel categories were evaluated separately (after normalization by rate of speech), and a multiple regression model was used to evaluate the performance of all vowel categories combined. Experiments on an English language proficiency assessment show that the non-native speech GMMs outperform the native speech GMMs, and that all 13 vowels have significant correlations with human scores when the non-native speech GMMs are used. The multiple regression combination obtained correlations with human scores of 0.71 when transcriptions were used to extract the vowel durations and 0.64 when the Automatic Speech Recognition (ASR) output was used. The experiments demonstrate that the vowel duration feature based on non-native speech GMMs is a useful predictor of L2 proficiency and is robust to different datasets and situations.