<p>Depression is among the most common and most damaging mental health disorders seen across the globe, with considerable impact on the afflicted, on society, and on functionality. It affects an individual’s emotions and physical well-being and thus calls for the creation of robust, automatic diagnostic tools. The recent advances in deep learning opened active areas of research in multimodal depression detection, exploiting the synergies obtained by combining different data modalities. This paper introduces a two-stream hybrid deep learning model that coherently fuses two multimodalities (viz., audio and text) through 2D-CNNs and BGRUs. In the audio stream, the algorithm extracts 2D features like MFCCs, chroma, tonnetz, log-Mel spectrogram, and spectral contrast, processed through 2D-CNN. In the text stream, word embeddings and word index sequences are fed into a BGRU network for contextual comprehension. The late fusion process combines the softmax probabilities from each stream to produce final predictions for binary and multi-level depression classification. The model was evaluated on the DAIC-WOZ depression dataset. Results indicated that the hybrid framework wins over the single-stream ones since it yielded an F1 score of 0.89, a recall of 0.94, and a precision of 0.86 in a multimodal setting. Clearly beyond what other existing models have achieved! These findings clearly suggest the effectiveness of the proposed model in the actual applications of automated depression diagnosis.</p>