AdaT-DFL: A Semi-Supervised Framework for Speech Depression Detection with Limited Reliable Labels

The lack of labeled data is a significant challenge in speech depression detection (SDD) due to the extensive subjective assessment required for accurate labeling. In this work, we propose AdaT-DFL, an innovative semi-supervised learning framework designed to improve depression detection with a limited number of labeled data and counts of unlabeled data. AdaT-DFL includes a learnable depression feature encoder, which extracts depression features and generates pseudo-labels. The Adaptive Threshold Screening algorithm enables the incorporation of high-confidence unlabeled data into the training process. By minimizing a combination of contrastive class aggregation loss and classification loss, AdaT-DFL achieve competitive classification performance in comparison to fully supervised learning while using 30% labeled data. We evaluate AdaT-DFL on the CMDC and EATD-corpus datasets. The results of experiments illustrate that AdaT-DFL enhances the ability of encoder to capture more comprehensive representation of the data space.