The ability to sequence genomes has far outstripped approaches for deciphering the information they encode. Here we present a suite of techniques, based on ribosome profiling (the deep sequencing of ribosome-protected mRNA fragments), to provide genome-wide maps of protein synthesis as well as a pulse-chase strategy for determining rates of translation elongation. We exploit the propensity of harringtonine to cause ribosomes to accumulate at sites of translation initiation together with a machine learning algorithm to define protein products systematically. Analysis of translation in mouse embryonic stem cells reveals thousands of strong pause sites and unannotated translation products. These include amino-terminal extensions and truncations and upstream open reading frames with regulatory potential, initiated at both AUG and non-AUG codons, whose translation changes after differentiation. We also define a class of short, polycistronic ribosome-associated coding RNAs (sprcRNAs) that encode small proteins. Our studies reveal an unanticipated complexity to mammalian proteomes. Display omitted ► Ribosome-profiling technique reveals complexity of mammalian proteome ► Many transcripts previously characterized as noncoding are in fact translated ► Translation proceeds at 5.6 codons per second and stalls at Pro-Pro-Glu motifs ► mESC differentiation involves global shifts in upstream translation A high-resolution look at mammalian translation reveals unanticipated diversity in the resulting proteome, including peptide products from putative noncoding RNAs.