NMR spectra are often acquired under conditions that produce data with low signal-to-noise ratios. The propagation of this noise through the analysis process leads to a garbage in, garbage out problem that obstructs the goal of automated, high-throughput NMR resonance assignment and structure determination. We use a standard NMR test protein, human ubiquitin, to show that substantial amounts of noise can be preferentially removed from NMR peak lists through automated processing that exploits known constraints on and relationships between NMR spectra. We further show that the resulting increase in signal-to-noise improves the performance of automated assignment programs.