Inertial sensors like seismometers or accelerometers are sensitive to tilt motions. In general, from pure acceleration measurements, it is not possible to separate the tilt acceleration from the translational ground acceleration. This can lead to severe misinterpretation of seismograms. Here, we present three different methods that can help solving this problem by correcting translational records for dynamic tilt induced by ground deformation with direct measurements of rotational motions: (1) a simple time‐domain method, (2) a frequency‐domain method proposed by Crawford and Webb (2000) using a coherence‐weighted transfer function between rotation and acceleration, and (3) an adapted frequency‐domain method that corrects only those parts of the spectrum with coherence between translational acceleration and rotation angle higher than 0.5. These three methods are discussed in three different experimental settings: (1) a reproducible and precisely known laboratory test using a high‐precision tilt table, (2) a synthetic test with a simulated volcanic very‐long‐period event, and (3) a real data set recorded during the 2018 Mt. Kīlauea caldera collapse. All the three test cases show severe influence of tilt motion on the acceleration measurements. The time‐domain method and the adapted frequency‐domain method show very similar performance in all three test cases. Those two methods are able to remove the tilt component reliably from the acceleration record.