Orca Call Detection using CNN and Spectrograms

Abstract

The number of whales and dolphins getting killed each year is increasing day by day and we are not away from the day when these whales would go extinct. The increase in the number of acoustic data captured by acoustic sensors has made detecting these calls possible using multiple Machine Learning and Deep Learning models. The project aims to develop a Convolution Neural Network (CNN) classifier that will automatically identify the calls made by killer whales and detect the particular pods to which they belong from given audio samples. Here, the audio event detection is treated as an image classification problem, where the image is a spectrogram that is calculated using discrete Fourier transforms. The reason for analyzing spectrograms is that different whales have different spectra (frequency decompositions) and time variations which can be evaluated by dissimilar patterns in spectrograms. There are two major steps that involve detection of calls. First, we classify the call using our CNN model. Second, we apply template matching to determine the start time and end time of the call along with the pod to which the Orca belongs.

Keywords: CNN, Template Matching, spectrograms