The method of seismic ambient noise cross-correlations (NCFs) has been demonstrated to be applicable for time offset measurement, especially in evaluating instrument clocks. Continuous recording of seismic ambient noise data makes it possible to analyze the performance of seismic instruments online. However, long-term cross-correlation and stacking calculations are required to obtain accurate travel time from ambient noise, which greatly reduces the time resolution of instrument performance detection. Therefore, we propose a travel time extraction method based on time–frequency analysis, which could obtain the travel time accurately even though the NCF has a low signal-to-noise ratio (SNR), and it is applied to the measurement of clock offsets in seismic instruments. The method combines S Transform and dictionary learning to improve the SNR of NCFs, and uses a peak extraction algorithm based on short-time Fourier transform to obtain accurate travel times. Additionally, the travel time drift of the station pair can be obtained by calculating the travel time difference between causal and acausal parts of NCFs. Using the data from ambient noise observations and the real data with time errors to verify that the proposed method can obtain accurate travel time for NCF with SNR lower than 6 and it can pick up a time offset as low as one sampling point, the detectable change is 0.046% of travel time, which is crucial for detecting weak changes in the performance of seismic instruments.