Study on Depression Detection Model with Self-Supervised Speech Representation: The Architecture of wav2vec2-LSTM-CNN

Depression is a common mental disorder with increasing global incidence rates. Traditional diagnostic methods rely on subjective assessments, which are time-consuming, costly, and lack objectivity. To address these challenges, this study proposes a depression detection model based on self-supervised speech representations—wav2vec2-LSTM-CNN hybrid architecture. The innovations of this study are reflected in the following three aspects:(1) Feature extraction innovation: The wav2vec2 self-supervised learning framework is introduced into depression speech detection, significantly enhancing the robustness of feature representation through pre-trained speech embeddings. (2) Model architecture innovation: An LSTM-CNN hybrid network is designed to combine the temporal modeling capabilities of LSTM networks with the local feature classification advantages of CNNs, achieving multi-scale feature fusionExperiments were conducted on the DAIC-WOZ dataset, and the results show that the proposed model achieved an F1-macro score of 0.75, significantly outperforming existing state-of-the-art models such as SpeechFormer++ (0.733) and DepAudioNet (0.6259). This study provides an objective and efficient auxiliary diagnostic tool for early depression screening with clear clinical translation value. Future work will focus on multimodal physiological signal fusion and lightweight deployment to further enhance the model's generalization capability and practical applicability.