Real electromyographic (EMG) noise remains a challenge to remove from electrocardiograms (ECGs) due to its spectrum overlaps with the ECG morphology. Strong smoothing approaches reduce noise but may distort important ECG features, emphasizing the need for denoising techniques that preserve signal morphology. We propose a hybrid ECG denoising framework based on parallel processing and soft fusion. The first branch targets low-frequency components to preserve P/T-wave and ST-segment, while the second branch tracks high-frequency components to preserve sharp QRS-complexes. A third QRS activity branch is used exclusively to effectively localize R peaks and construct smoothly varying fusion weight through temporal smoothing, enabling gradual transitions and avoiding switching artifacts. The method is evaluated on the SimEMG database, which provides EMG-contaminated ECG recordings. Mean SNR improvement was 8.27, 7.72, 5.32, 3.41, and 2.20dB for input SNR <4, 4-8, 8-12, 12-16, 16-20dB, respectively. Average correlation increased up to 0.995 with PRD values decreased to reach 10.25%. EMG noise can be effectively suppressed by using soft mask fusion while preserving ECG morphology without needing reference data, reflecting a robust performance to be a strong solution for real-world noise removal.