Best way to normalize them depends on your goal and the material you study. For glasses and crystals, one useful way is however to normalise Raman intensities by the total area of the Raman spectrum. This area will thus be equal to 1 after normalisation, and it will be consistent with the sum of the density of state of your glass/crystal that is equal to 1 in theory.
It depends on what goal you want to achieve. You can left it as is for presentation (naming it Raman intensity, arbitrary units), you can recalculate it to Raman activity to compare with theoretical calculation, or normalize largest peak or integral between peaks and baseline for machine learning based classification.
Say if i want to compare raman shift and FWHM change of samples (under process optimisation) against reference monocrystalline silicon baseline, what would be the best way? I often see people normalised the intensity in publications. Thanks a lot guys.
Normalization to the intensity of the highest peak is certainly the easiest approach, and will likely give you the answers you want in this case, but I agree with Charles that the most robust normalization is to the total integrated area of the spectrum The problem with normalizing to intensity alone arises when a processing condition causes a dramatic change in the most intense feature, or if a feature at a different frequency becomes stronger than the initial feature. These would cause very strange-looking results in your analysis, as the initial normalization condition no longer holds. Usually, with silicon spectra, there is little chance of overwhelming the 520 cm-1 feature, so intensity normalization is a quick and easy solution in this case.