The peripheral oxygen saturation (SpO2) reflects the metabolic capacity of the human body, which can be used in assessing or screening symptoms such as pulmonary embolism. Photoplethysmography (PPG) is a common method for SpO2 monitoring, while it suffers from interferences such as ambient light scattering, reflections, and motion artifacts. These interferences significantly degrade the accuracy of SpO2 monitoring. Filtering techniques are widely used to suppress the interferences in PPG signals. However, conventional PPG filtering techniques use a fixed window, which is not able to handle interferences at different frequencies. In this paper, we propose a flexible-window filtering technique to suppress the interferences in SpO2 monitoring. To validate the proposed technique, we built a prototype to monitor the in-vivo SpO2 of the human body. Measurement results show that the proposed technique reduces the mean absolute percentage error (MAPE) of SpO2 by 46% compared to the conventional methods.