In the last years, the geophysics community has been devoted to seismic data quality enhancement by noise attenuation and seismogram interpolation using CNN-based methods. Discriminative CNN-based methods can achieve state-of-the-art denoising results. However, they do not apply to scenarios without paired training data (i.e., noisy seismic data and corresponding ground-truth noise-free seismic data). In this work, we treat seismic data denoising as a blind denoising problem to remove unknown noise from noisy shot gathers without ground truth training data. The basis used by the denoiser model is learned from the noisy samples themselves during training. Motivated by this context, the main goal of this work is to propose a selfsupervised method for blind denoising of seismic data, which requires no prior seismic signal analysis, no estimate of the noise, and no paired training data. Our proposed self-supervised method assumes two given datasets: one containing noisy shot gathers and the other noise-free shot gathers. From this data, we train two models: (1) Seismic Noise Transfer (SNT), which learns to produce synthetic-noisy shot gathers containing the noise from noisy shot gathers and the signal from noise-free shot gathers; And (2) Seismic Neural Denoiser (SND), which learns to map the syntheticnoisy shot gather back to original noise-free shot gather. After training, SND alone is used to remove the noise from the original noisy shot gathers. Our SNT model adapts the Neural Style Transfer (NST) algorithm to the seismic domain. In addition, our SND model consists of a novel multi-scale feature-fusion-based CNN architecture for seismic shot gather denoising. Our method produced promising results in a holdout experiment, achieving a PSNR gain of 0.9 compared to other state-of-the-art models.