Extracting dolphin whistles in complex acoustic scenarios: a case study in the Bay of Biscay

Accurate whistle contour extraction is crucial in many dolphin behavioural studies. Traditionally, whistle contour extraction involves a first step of finding whistle candidates by peak-level detection in the time-frequency domain, followed by a determination of when peaks are close enough to each other to be part of the same whistle contour. In complex scenarios, such as those with a large number of individuals vocalising simultaneously or those with a sudden increase in background noise, peak-level detection may not provide a number of accurate whistle candidates that is large enough to extract the whistle contour or to disambiguate individual whistles when they cross one another. In these adverse scenarios, a different approach, based on the pyknogram representation, can produce a more accurate detection of whistle candidates and evenly distributed candidates throughout the duration of the whistle. This work compares the peak-level extraction approach of the spectrogram with the point-density extraction approach of the pyknogram. We propose a technique that combines estimates of the central frequency and bandwidth to extract whistle candidates in adverse scenarios. The method has been successfully used for the vocalisation extraction of dolphins in the Bay of Biscay (Spain) using a database of more than 2000 dolphin whistles.