The main limitation of MRS is low SNR. Several approaches for denoising have been proposed. However, it is debatable, whether denoising can reduce estimate uncertainties. In this work, we investigate denoising using deep learning (DL) in time-frequency representations. The results were assessed using two methods: first, an adjusted noise score was used and second, the outcome of traditional fitting was evaluated. We found that time-frequency domain denoising through DL produces a visually appealing spectrum but mean residuals for the relevant spectral regions and variance in fit results remained as high as without denoising.