Depression detection methods based on multimodal fusion of voice and text

Depression is a prevalent mental health disorder, and early detection is crucial for timely intervention. Traditional diagnostics often rely on subjective judgments, leading to variability and inefficiency. This study proposes a fusion model for automated depression detection, leveraging bimodal data from voice and text. Wav2Vec 2.0 and BERT pre-trained models were utilized for feature extraction, while a multi-scale convolutional layer and Bi-LSTM network were employed for feature fusion and classification. Adaptive pooling was used to integrate features, enabling simultaneous depression classification and PHQ-8 severity estimation within a unified system.Experiments on the CMDC and DAIC datasets demonstrate the model's effectiveness. On CMDC, the F1 score improved by 0.0103 and 0.2017 compared to voice-only and text-only models, respectively, while RMSE decreased by 0.5186. On DAIC, the F1 score increased by 0.0645 and 0.2589, with RMSE reduced by 1.9901. These results highlight the proposed method's ability to capture and integrate multi-level information across modalities, significantly improving the accuracy and reliability of automated depression detection and severity prediction.