Ultrasonic fatigue testing (USFT) is a time-efficient method for evaluation of the fatigue limit of metallic alloys in a high and very high cycle range. Propagation of a final crack at the very end of the fatigue life can be clearly detected by USFT parameters and monitoring techniques. In contrast, initiation of fatigue damage during USFT remains unclear. Despite its excellent sensitivity, the implementation of acoustic emission (AE) for damage monitoring is hindered by severe noisiness of acoustic signals from USFT. In addition to resonance-related quasi-stationary noises, USFT in a pulse-pause mode is accompanied by non-stationary ones, which makes recognition of material-related AE signals even more difficult. A special AE processing algorithm was developed to overcome this issue and get useful insights from USFT monitoring. Consistent cropping, Fourier transformation, adaptive filtration and thresholding allowed to calculate noise-free AE activity during USFT of 42CrMo4 steel. Remarkably, most of this activity was located in the very beginning of fatigue loading. In comparison to non-failed runout samples, AE activity of samples with fatigue cracks was significantly higher, indicating its relation to fatigue damage. The proposed AE processing principle can be helpful for monitoring of USFT samples and other parts, operating in resonance conditions.