Styrian dialect classification: comparing and fusing classifiers based on a feature selection using a genetic algorithm

Many classifiers struggle when confronted with a high dimensional feature space like in the data sets provided for the Interspeech ComParE challenge. This is because most features do not significantly contribute to the prediction. To alleviate this problem, we propose a feature selection based on a Genetic Algorithm (GA) that uses an SVM as the fitness function. We show that this yields a reduced subset (1) which results in an Unweighted Average Recall (UAR) that beats the challenge baseline on the development set for the 3-class classification problem. Further, we extract an additional per-phoneme feature set, where the features are inspired by the ComParE features. On this set the same GA-based feature selection is performed and the resulting set is used for training in isolation (2) and in combination with the aforementioned reduced challenge features (3). Five classifiers were tested on the three subsets, namely SVMs, DNNs, GBMs, RFs, and regularized regression. All classifiers achieved a UAR above the baseline on all three sets. The best performance on set (1) was achieved by an SVM using an RBF kernel and on sets (2) and (3) by a fusion of classifiers.