A Novel Tracker of Adaptive Directional Ridge Separation and Prediction for Detecting Whistles

Abstract

Whistle detection of marine mammal signals with close and overlapping components of varying amplitudes is a key task for overlapping source separation. In this article, we propose a novel tracker, called adaptive directional ridge separation and prediction, for detecting whistles, which are typically analyzed using a time-frequency (TF) representation. Inspired by TF reassignment, a new reassignment scheme based on time-scale changes is developed to acquire instantaneous TF points with high energy concentration. To address the mutual interference among various types of components, a tone-pulse separation model is introduced for the aliased TF components, utilizing these instantaneous TF points and instantaneous rotating operators. An adaptive directional ridge predictor is established for application in automatic overlapping whistle detection, ensuring unbroken detection even when a whistle becomes nearly indistinguishable in the TF representation. Experimental results, obtained using both a simulated signal and recorded calls of marine mammals, demonstrate the superiority of the proposed method compared to other state-of-the-art methods. This method is capable of performing whistle detection and separating overlapping sources even in the presence of splash noises, which may cause partial distortion or disconnection of components from the TF representation.