Abstract We develop a methodology to compile an objective tremor catalog by utilizing distinctive event features that differentiate tectonic tremors from non‐tremor events, and combining the envelope cross‐correlation method with clustering technique and neural network. This approach enables tremor extraction without subjective criteria, allowing for the detection of previously overlooked short‐duration tremors. The event features employed to distinguish tremors and non‐tremor events are depth, the mean amplitudes at high and low frequencies, the ratio of these two amplitudes, and event duration. The duration is defined as the minimum period that contains 50% of the seismic energy. The application of this method to western Japan detects 1.7 times more tremors than the previous studies, with the durations of 0.3–∼100 s. The events with short durations are considered low‐frequency earthquakes. The relationship between seismic moment and duration of the detected tremors is consistent with the scaling law of slow earthquakes.