Intelligent identification of the Iraqi Arabic accent within operational and security environments, often characterized by telecommunication and radio channel noise, constitutes a critical challenge in speech processing. Conventional accent detection models, which rely primarily on spectral features (like Mel-Spectrograms), often fail to capture vital phase information and temporal fine structures of the signal at low Signal-to-Noise Ratios (SNR). To address this limitation, this paper proposes a novel hybrid architecture that leverages the synergy between the Whisper Large Language Model (LLM) and the SincNet neural network. In the proposed framework, Whisper is utilized to extract robust semantic and high-level spectral embeddings, while SincNet processes the raw waveform directly to extract subtle cues related to speaker identity and temporal patterns. The primary novelty of this research lies in the employment of an “Attention-based Feature Fusion” mechanism. This module dynamically and intelligently weights the significance of features extracted from both branches, thereby effectively mitigating the adverse effects of noise. Experimental evaluation on the standard ADI17 benchmark dataset under simulated radio noise conditions (SNR=10 dB) demonstrates that the proposed method achieves a superior accuracy of 93.53%. This result not only significantly improves upon the detection accuracy of standalone Whisper-based baselines but also surpasses state-of-the-art deep learning approaches. The findings substantiate the efficacy and robustness of this hybrid architecture for deployment in sensitive operational and security applications.