Automated depression detection (ADD) from speech signals allows early identification and intervention, reducing costs to medical healthcare. However, most of the existing ADD studies are trained and evaluated on a single language corpus with a lack of sufficient training data. These limits the generalizability of models in other demographic groups in distinct languages. In this study, Semi-Supervised Learning (SSL) was applied to depression detection on two different language datasets. We evaluate the HuBERT and WavLM models in single-language, mixed-language, and cross-language scenarios to investigate the generalization to diverse populations at different recording environments. Moreover, we thoroughly analyzed layer-wise performance in the upstream model and pooling methods (i.e. max and mean pooling) in the downstream task. The results show that the WavLM features generalize better than the HuBERT features. Our best model surpasses previous works in the frozen upstream conditions.