Depression detection from speech has emerged to be the most advanced area in the field of medical analysis. The depression in patients often leads to a wide range of mental issues along with suicidal tendencies. So in this research, we propose a hybrid framework combining a fusion of convolutional and recurrent representations along with the handcrafted acoustic features in an MLP pipeline for consistent binary classification of depressed and not depressed classes. In this research, this hybrid model was trained using the depression dataset DAIC-WOZ clinical interview corpus and tested on the emotion dataset RAVDESS by mapping the emotions into both negative and positive classes, to assess cross-corpus generalization, which is an essential requirement for real-world implementation. The proposed CNN–BiLSTM fusion architecture extracts high-level temporal–spectral embeddings from Mel-spectrograms and integrates them with statistical prosodic descriptors such as spectral contrast, MFCCs, chroma, and energy-based measures. Integration of augmentation techniques to handle data imbalance is applied to enrich minority depressive samples. Experimental results showcase that the proposed model achieves an accuracy of 87.4% on the diac-woz dataset with the F1-score, precision, and recall of 87.4%, 86.8% and 87.9%, respectively. Along with on cross corpus evaluation with the RAVDESS dataset accuracy obtained is around 80.014%. These findings highlight the possibility of integrating deep temporal modeling and handcrafted feature fusion for generalized depression detection across diverse speech datasets, advancing the field toward effective and trustworthy affective computing systems.