Loud transients introduce bias to background spectrum estimates based on the sample mean, like Welch's Overlapped Segment Averaging (WOSA). The Schwock and Abadi Welch Percentile (SAWP) estimator avoids the loud transient bias by replacing the averaging of the WOSA with a scaled order statistic (OS) of the sample power spectrum. However, choosing the percentile for the SAWP is a challenging problem. SAWP estimators based on lower percentiles are more robust against loud transients, but also have a higher variance. Also, the rate of occurrence of loud transients may change with time, requiring the SAWP to adapt which percentile is used. To approach this challenge, this work proposes a Universal SAWP estimator which is a weighted sum across different fixed percentile SAWP estimators. At each iteration, the Universal SAWP uses techniques from Singer and Feder's universal linear prediction to update the blend weights, promoting the percentiles with the lowest sample variance over recent observations. In computer simulations, the Universal SAWP quickly assigns higher weights to the lowest variance estimators as the occurrence of loud transients increases. Overall, the Universal SAWP achieves lower variance and mean squared error than the best fixed percentile SAWP estimator.

This content is only available via PDF.