<p>Real-world speech enhancement remains challenging due to highly non-stationary noise, multi-scale spectral variations, and phase distortion in noisy recordings. To address these challenges, this paper proposes a complex-domain speech enhancement framework that jointly models magnitude and phase through a unified architectural and optimization strategy. The proposed method operates on compressed complex spectrograms and integrates a Dilated DenseNet encoder for multi-scale local feature extraction with a Dual-Path Transformer architecture incorporating two-stage conformer blocks to capture both local and global temporal-spectral dependencies. A dual-branch decoder is employed to decouple magnitude mask estimation from complex residual refinement, enabling coherent reconstruction of amplitude and phase. To further enhance perceptual quality, training is guided by a composite loss function that combines magnitude and real-imaginary regression with perceptually motivated cepstral and multi-resolution Short-Time Fourier Transform (STFT) losses, complemented by time-domain supervision. Experimental evaluations on the VoiceBank+DEMAND dataset, which is a widely used speech enhancement benchmark combining the VoiceBank clean speech corpus with the DEMAND environmental noise recordings. The proposed framework achieves a Perceptual Evaluation of Speech Quality (PESQ) score of 2.75, Short-Time Objective Intelligibility (STOI) of 0.93, and Segmental Signal‑to‑Noise Ratio (SSNR) of 9.58 dB, outperforming recent state-of-the-art models, particularly under low‑signal‑to‑noise‑ratio (SNR), multi-source, and reverberant noise conditions. These results confirm that jointly leveraging hierarchical temporal-spectral modeling and perceptually aligned loss functions leads to substantial improvements in both signal fidelity and perceptual speech quality.</p>