Mental health is a global issue that plays an important roll in the overall well-being of a person. Because of this, it is important to preserve it, and conversational systems have proven to be helpful in this task. This research is framed in the MENHIR project, which aims at developing a conversational system for emotional well-being monitorization. As a first step for achieving this purpose, the goal of this paper is to select the features that can be helpful for training a model that aims to detect if a patient suffers from a mental illness. For that, we will use transcriptions extracted from conversational information gathered from people with different mental health conditions to create a data set. After the feature selection, the constructed data set will be fed to supervised learning algorithms and their performance will be evaluated. Concretely we will work with random forests, neural networks and BERT.

Concern for mental health has increased in the last years due to its impact in people life quality and its consequential effect on healthcare systems. Automatic systems that can help in the diagnosis, symptom monitoring, alarm generation etc. are an emerging technology that has provided several challenges to the scientific community. The goal of this work is to design a system capable of distinguishing between healthy and depressed and/or anxious subjects, in a realistic environment, using their speech. The system is based on efficient representations of acoustic signals and text representations extracted within the self-supervised paradigm. Considering the good results achieved by using acoustic signals, another set of experiments was carried out in order to detect the specific illness. An analysis of the emotional information and its impact in the presented task is also tackled as an additional contribution.

Emotion recognition from speech is an active field of study that can help build more natural human-machine interaction systems. Even though the advancement of deep learning technology has brought improvements in this task, it is still a very challenging field. For instance, when considering real life scenarios, things such as tendency toward neutrality or the ambiguous definition of emotion can make labeling a difficult task causing the data-set to be severally imbalanced and not very representative.In this work we considered a real life scenario to carry out a series of emotion classification experiments. Specifically, we worked with a labeled corpus consisting of a set of audios from Spanish TV debates and their respective transcriptions. First, an analysis of the emotional information within the corpus was conducted. Then different data representations were analyzed as to choose the best one for our task; Spectrograms and UniSpeech-SAT were used for audio representation and Dis-tilBERT for text representation. As a final step, Multimodal Machine Learning was used with the aim of improving the obtained classification results by combining acoustic and textual information.