Voice pathology assessment using sustained vowels has proven to be effective and reliable. However, only a few studies regarding detection of pathological speech based on continuous speech are available. In this study we evaluate the usefulness of various regression models trained on continuous speech recordings from Saarbruecken Voice Database in the detection of voice pathologies. The recordings were used for extraction of speaker embeddings called x-vectors based on mel-frequency cepstral coefficients and gammatone frequency cepstral coefficients. Since the dataset used in this study is imbalanced, various over- and undersampling techniques were applied to the training set to ensure robustness of models’ decision boundaries. The models were trained on both imbalanced and resampled training sets using 5-fold cross-validation. The best results were obtained for Multi Layer Perceptron trained on GFCC-based x-vectors, achieving accuracy of 0.8184, F1-score of 0.8212, and ROC AUC score of 0.8810 for the testing set.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.