Spectral rendering approaches are currently the most accurate ones for predictively simulating the light transport equation with advanced wavelength-dependent phenomena such as polarization, light dispersion, or metamerism. However, these approaches require a much larger number of samples than standard trichromatic approaches - e.g. expressed in RGB or HSV color spaces - to properly cover the visible spectral domain and then achieve noise-free images without chromatic aberration. To address this issue, denoising solutions can be considered and applied to a spectral renderer, while still ensuring that such an algorithm can properly handle wavelength-dependent features and induced potential color noise. However, to our knowledge, no spectral rendering denoiser exists in state-of-the-art. In this paper, we propose one inspired by a feature fusion based on an aggregation of a deep dual-encoder network with a deep combiner. We use a deep-learning (DL) approach that uses Principal Component Analysis (PCA) transformation to compress all the spectral wavebands into a single parametric image. This parametric feature is then added as an auxiliary input to material information. It is encoded with normal, depth and spectral albedo to take into account the chromatic specificities of the noise induced by spectral rendering. Our experiments demonstrate that our method outperforms state-of-the-art denoising techniques for spectral rendering, while significantly reducing the number of samples required for noise-free images. It provides a better reconstruction than the RGB and XYZ approaches, while preventing the appearance of most artifacts.