Automatic Native Language Identification Using Novel Acoustic and Prosodic Feature Selection Strategies

Abstract

We consider the problem of automatic identification of native language (L1) of non-native English (L2) speakers from eleven L1 backgrounds. Analyzing the influence of each L1 pronunciation variabilities on L2 pronunciation, different sets of linguistic units are chosen to compute supra-segmental features by considering the acoustic and prosodic variations within and across these sets. Using these features, we implement a multi-class classifier comprising 55 binary (one versus another) support vector machine (SVM) classifiers. We select optimal set of features for each binary classifier using two feature selection strategies (FSSs) based on Fisher discriminant ratio (FDR). The first strategy considers the features that maximizes the each binary classifier performance. However, the second strategy selects the features by maximizing a multi-class classifier performance for which an algorithm is proposed. Experiments are performed on the ETS corpus of non-native spoken English, comprising 4099 files. When the proposed features along with FSSs are used, the unweighted average recall (UAR) on the test set for each selection strategy is found to be 1.3% and 2.1% (absolute) higher compared to using all features; as well as 3.0% and 3.8% higher than the baseline technique respectively.