Depression is a frequent and curable psychiatric disorder, detrimentally affecting daily activities, harming both work-place productivity and personal relationships. Among many other symptoms, depression is associated with disordered speech production, which might permit its automatic screening by means of the speech of the subject. However, the choice of actual features extracted from the recordings is not trivial. In this study, we employ x-vectors, a DNN-based feature extractor technique, to detect depression from a Hungarian corpus. We experiment with training custom x-vector extractors, and we also explore the performance of an out-of-domain pre-trained one. Our findings confirm that x-vectors are able to capture meaningful speaker traits that contain information for depression discrimination. We also show that the language of the extractor is of secondary importance compared to the frame-level feature set: our best model, which achieved an AUC score of 0.940 and an RMSE score of 9.54, was trained on log-energies instead of MFCCs.

Dementia is a chronic or progressive clinical syndrome, mainly characterized by the deterioration of memory, thinking, reasoning and language. In Mild cognitive impairment (MCI), often considered as the prodromal stage of dementia, there is also a subtle deterioration of these functions, but they do not affect the daily life of the patient. However, due to the slight nature of the changes, it is quite hard to diagnose MCI. In this study, we employ sequence-to-sequence deep autoencoders in order to extract compact, robust and efficient attributes from the spontaneous speech of 25 MCI subjects and 25 healthy controls. From our results, this approach gives a competitive performance, as we significantly outperformed x-vectors even though they were trained on more data. Our additional efforts to identify mild Alzheimer’s (mAD) subjects as well were less successful; but since the focus is on the early detection of dementia, this is not a limitation of the methodology from a practical point of view.

Estimating the degree of sleepiness from the human speech is an emerging research problem with straightforward applications. In this study, we employ the x-vector approach, currently the state-of-the-art in speaker recognition, as a neural network feature extractor to detect the level of sleepiness of a speaker. Besides using different corpora for fitting the x- vector DNN, we also experiment with adding noise and reverberation to the training samples. According to our experimental results for the publicly available Dusseldorf Sleepy Language Corpus, utilizing x-vector embeddings as features for Support Vector Regression consistently leads to competitive performance scores in sleepiness detection. In particular, we present the highest Spearman's correlation coefficient on the public corpus that was achieved by a single method.

Computational paralinguistics is concerned with the automatic identification of non-verbal information in human speech.The Interspeech ComParE challenge features new paralinguistic tasks each year; this time, among others, a cross-corpus conflict escalation task and the identification of primates based solely on audio are the actual problems set.In our entry to ComParE 2021, we utilize x-vectors and Fisher vectors as features.To improve the robustness of the predictions, we also experiment with building an ensemble of classifiers from the x-vectors.Lastly, we exploit the fact that the Escalation Sub-Challenge is a conflict detection task, and incorporate the SSPNet Conflict Corpus in our training workflow.Using these approaches, at the time of writing, we had already surpassed the official Challenge baselines on both tasks, which demonstrates the efficiency of the employed techniques.

Parkinson's Disease (PD) is a neuro-degenerative disorder that affects primarily the motor system of the body.Besides other functions, the subject's speech also deteriorates during the disease, which allows for a non-invasive way of automatic screening.In this study, we represent the utterances of subjects having PD and those of healthy controls by means of the Fisher Vector approach.This technique is very common in the area of image recognition, where it provides a representation of the local image descriptors via frequency and high order statistics.In the present work, we used four frame-level feature sets as the input of the FV method, and applied (linear) Support Vector Machines (SVM) for classifying the speech of subjects.We found that our approach offers superior performance compared to classification based on the i-vector and cosine distance approach, and it also provides an efficient combination of machine learning models trained on different feature sets or on different speaker tasks.