A theoretical triglyceride model was developed for in vivo human liver fat ¹H MRS characterization, using the number of double bonds (CHCH), number of methylene‐interrupted double bonds (CHCHCH₂CHCH) and average fatty acid chain length. Five 3 T, single‐voxel, stimulated echo acquisition mode spectra (STEAM) were acquired consecutively at progressively longer TEs in a fat–water emulsion phantom and in 121 human subjects with known or suspected nonalcoholic fatty liver disease. T₂‐corrected peak areas were calculated. Phantom data were used to validate the model. Human data were used in the model to determine the complete liver fat spectrum. In the fat–water emulsion phantom, the spectrum predicted by the model (based on known fatty acid chain distribution) agreed closely with spectroscopic measurement. In human subjects, areas of CH₂ peaks at 2.1 and 1.3 ppm were linearly correlated (slope, 0.172; r = 0.991), as were the 0.9 ppm CH₃ and 1.3 ppm CH₂ peaks (slope, 0.125; r = 0.989). The 2.75 ppm CH₂ peak represented 0.6% of the total fat signal in high‐liver‐fat subjects. These values predict that 8.6% of the total fat signal overlies the water peak. The triglyceride model can characterize human liver fat spectra. This allows more accurate determination of liver fat fraction from MRI and MRS. Copyright © 2010 John Wiley & Sons, Ltd.