Depression is one of the most common mental health disorders, affecting approximately 332 million people worldwide. The social and health impact of this condition is severe, with suicide being the third leading cause of death among people aged 15 to 29 [55]. Currently, diagnosing depression relies primarily on patient self-reporting questionnaires and interviewbased clinical assessments, with the Patient Health Questionnaire (PHQ) [71] being the most common for diagnosis and direct observation. However, technological advances have driven strategies for identifying depression, which explore the detection of depressive symptoms through the analysis of multimodal data, such as text and voice. Extensive research on this topic has been conducted, mainly on video data, text, and voice. Supported by medical research indicating that depression is associated with impaired emotional regulation, we propose an alternative approach based on the idea that estimating the level of depression may benefit from the simultaneous learning of emotion intensity. Therefore, we hypothesize that the integration of audio and textual features with emotion-related information extracted from both modalities may improve the performance of voice-based depression detection systems and provide a novel approach to support clinical decision-making. Furthermore, such systems could contribute to the early identification of severe depressive episodes, enabling timely adjustments to treatment and medication strategies. Deep learning approaches applied to textual and audio data have demonstrated promising performance in depression detection tasks. However, several limitations have been identified. First, the availability of sufficiently large datasets remains limited, as most existing datasets are relatively small in size. In addition, certain proposed architectures require extensive preprocessing procedures, such as the extraction of handcrafted features from raw audio signals to achieve improved performance. Furthermore, some architectures present limitations in effectively capturing long-range dependencies within sequential data. The aim of this project is to fine-tune several transformer-based models for textual and audio data, as well as to employ pretrained transformer models for emotion extraction. These systems are evaluated and subsequently integrated into a meta-learning framework designed to classify patients as depressed or non-depressed in a binary classification context. In addition, an interpretability analysis based on SHAP is conducted to identify and explain the features that contribute most significantly to the meta-model’s predictions..