Abstract
When estimating the background noise power spectral density (PSD) from underwater acoustic recordings, order statistics filters (OSF) effectively mitigate the bias caused by outliers in data, such as broadband loud transients. The Schwock and Abadi [ICASSP, 2019] Welch Percentile (SAWP) is an example of a spectral estimator that scales a single-order statistic (OS) of consecutive overlapping periodograms of acoustic data to estimate the background noise PSD. However, in a dynamic environment, the rate at which loud transients occur is time-varying, requiring the OSF to adjust its rank accordingly to keep low bias and variance. Previously, we proposed applying a mixture of experts to blend SAWP estimators of different ranks according to their short-time performance, thus eliminating the need to explicitly set the OSF rank [Campos Anchieta & Buck, POMA, 2024]. The performance of each estimator is measured by their sample variance over a fixed time window. In this talk, we apply the same performance-weighted blend (PWB) algorithm to the truncated linear order statistics filter (TLOSF), an OSF that is itself a weighted sum of OS up to a threshold rank [Campos Anchieta & Buck, IEEE JOE, 2024]. When compared to any of their fixed rank counterparts, the PWB versions of both SAWP and TLOSF accumulate less squared error estimating the PSD. When compared to each other, the performance-weighted TLOSF has 0.25–0.5 dB lower mean squared error than the PWB SAWP mainly due to a lower variance. [Work supported by ONR Code 321US.]