Since the train and test partitions in these works differ slightly from our experiments, an exact comparison is not feasible.
Performances could probably be increased by choosing a smaller test partition and thus increasing the training material.
Each cohort is used as a test partition, and the remaining subsets are combined to complete the function.
For each fold, we first normalized the frame-level features of the test partition to have the same mean as that of the training partition.
Cross-validation was not required for this database, since the training, development, and test partitions, as specified in [18,47], do not overlap in the speakers.
Results, presented in Table 4, are the average of 20 runs using randomly generated training, validation and test partitions for each run.
This produces 10 models that are all trained on different subsets of the original dataset and where each has been used as the test partition one time only.
For testing embedding, each network was trained for 100 epochs with experiments repeated 20 times to eliminate bias, due to chance split when creating training, validation and test partitions.
Instead, we trained a new classifier for Pre-diagnosis, using training and test partitions contained within the Pre-diagnosis data, which left the Pre-diagnosis model with considerably fewer data points to train on.
Specifically, we used the full ECBDL14 dataset to show in Table 6 the true rates for the positive and negative classes obtained by the decision tree in the test partitions.
