Challenges of using longitudinal and cross-domain corpora on studies of pathological speech

Several promising works have reported very exciting results in the field of speech in health, however there are still issues to address before deploying such systems into clinical applications. One of such issues is to ensure the generalisability and reliability of results. With this in mind, in this work, we perform a comparative analysis of healthy speech in two scenarios: (1) collected for six different datasets spoken in the same language, and (2) collected across different times in a single longitudinal corpus. We show that feature sets typically used for disease detection from speech (eGeMAPS, ComParE, ECAPA-TDNN embeddings and i-vectors) encode much information about the dataset or about changing recording conditions over time, in longitudinal studies. We support our results with classification results largely above chance level for both scenarios, and through unsupervised clustering experiments, where we observe that data naturally clusters according to dataset.