Bioacoustic transient detection by image convolution

A method is presented for detecting bioacoustic transient signals. The method detects a common feature of animal calls, frequency sweeps, or combinations of frequency sweeps over time, by convolving a spectrogram image with a two‐dimensional kernel designed to find lines in images. This kernel has positive (excitatory) regions to achieve strong convolution operator response on desired frequency sweeps, and negative (inhibitory) regions to reject noise and interfering sounds. Design parameters control permissible variation in the sweeps. This technique is useful in screening long recordings for desired call types and perhaps for identifying individuals. For screening, it is compared with two other approaches on recordings of Bowhead Whale calls in a noisy Arctic environment: a hidden Markov model and a matched filter. This method performs better than both others on the Bowhead data, perhaps well enough for practical use. It also detects perfectly 15‐ to 20‐Hz calls of unknown origin, possibly blue whales, ...