Functional MRS (fMRS) is a powerful technique to measure metabolite responses over time. However, noise and spectral contamination limit the ability to study individual metabolite time-courses. In this work, we propose to model fMRS spectra as a superposition of low-rank (L) and sparse components (S). L+S decomposition resulted in separation of temporally-correlated signal from noise in simulation. In vivo, L+S spectra had higher SNR compared to original data (P=0.007) and the mean glutamate time-course, using L+S spectra, was more strongly correlated to stimulus. L+S decomposition is a promising data-driven method to enhance sensitivity to dynamic changes in fMRS.