Predicting a Cold from Speech Using Fisher Vectors; SVM and XGBoost as Classifiers

Screening a cold may be beneficial in the sense of avoiding the propagation of it. In this study, we present a technique for classifying subjects having a cold by using their speech. In order to achieve this goal, we make use of frame-level representations of the recordings of the subjects. Such representations are exploited by a generative Gaussian Mixture Model (GMM) which consequently produces a fixed-length encoding, i.e. Fisher vectors, based on the Fisher Vector (FV) approach. Afterward, we compare the classification performance of the two algorithms: a linear kernel SVM and a XGBoost Classifier. Due to the data sets having a high class imbalance, we undersample the majority class. Applying Power Normalization (PN) and Principal Component Analysis (PCA) on the FV features proved effective at improving the classification score: SVM achieved a final score of 67.81% of Unweighted Average Recall (UAR) on the test set. However, XGBoost gave better results on the test set by just using raw Fisher vectors; and with this combination we achieved a UAR score of 70.43%. The latter classification approach outperformed the original (non-fused) baseline score given in ‘The INTERSPEECH 2017 Computational Paralinguistics Challenge’.