Abstract: The spectral ratio method is a technique commonly used for estimating Q values. It calculates the quality factor Q based on the linear relationship between the logarithmic spectral ratio of seismic wavelets at different times and the Q value. The core idea of the time-frequency domain spectral ratio method is to extract the spectral distribution at different moments via time-frequency spectrum analysis, representing the frequency spectrum of seismic wavelets at those moments while ignoring the influence of reflection coefficients. However, because of the impact of local reflection coefficients, the logarithmic spectral ratios of seismic wavelets obtained via the time-frequency spectral ratio method often exhibit complex multi-peak patterns, deviating from the linear relationship between logarithmic spectral ratios and frequency. This leads to significant errors in Q values derived from linear fitting, which fail to meet the requirements for lithological interpretation and pore fluid prediction. To address the issues, the present study combines cepstral analysis with the Generalized S-transform to propose a time-varying wavelet spectrum extraction method that eliminates the influence of reflection coefficients, thereby improving the Q estimation accuracy. First, the Generalized S-transform is applied to convert seismic records into the time-frequency domain. The logarithmic time-frequency spectrum of the seismic records is then computed to obtain their cepstra at different times. In the cepstral domain, the wavelet components reside in the low-frequency band, whereas reflection coefficients occupy the mid-to-high-frequency bands. By applying least-squares smoothing to the low-frequency cepstral components, the wavelet cepstrum is derived. Taking the exponential of the cepstrum reconstructs the spectrum of the time-varying wavelet. Finally, Q values are calculated using the linear relationship between the logarithmic spectral ratios of adjacent wavelets and the Q value. Model tests validate the accuracy of time-varying wavelet extraction, as well as the feasibility and noise resistance of Q estimation. Application to real seismic data from an oilfield in eastern China further confirms the effectiveness of the method.