Depression is a prevalent and disabling mental health condition that affects over 300 million people worldwide, yet it remains significantly underdiagnosed due to the reliance on self-reports and clinical interviews. Accurate and non-invasive detection of depression severity from speech has the potential to enhance large-scale mental health screening. As an alternative, speech has emerged as a promising non-invasive biomarker for depression assessment, reflecting subtle changes in prosody, rhythm, and vocal expression. In this work, we propose a dual-stream Conformer-based acoustic modeling framework to predict depression severity scores directly from speech descriptors. Specifically, we leverage two complementary acoustic feature sets—MFCC and eGeMAPS—which are processed through a dual-branch Conformer architecture to capture both spectral and prosodic patterns in speech. The outputs from both streams are then fused and transformed into a unified representation, which is used to estimate depression severity through a regression layer. To aggregate temporal dynamics, we employ self-attention pooling to extract a compact utterance-level representation, which is subsequently passed through a regression head to estimate the depression score. We evaluate our approach on a clinically annotated dataset containing depression severity ratings. The proposed model achieves a root mean square error (RMSE) of 4.87 and a mean absolute error (MAE) of 4.68 on the depression score regression task. In addition, binary classification of depressive vs. non-depressive speech based on predicted scores reaches an accuracy of 74.91%, demonstrating the effectiveness of the proposed acoustic fusion architecture for both regression and screening purposes.