A two-stage process for automatic recognition of harbor seal vocalizations

Abstract

Male harbor seals (Phoca vitulina) were continuously recorded for a year from an array of hydrophones in shallow water off the coast of central California. A two-stage automatic recognition system was used to extract sounds of interest. The first stage, fast but crude, processed the entire sound archive. It operated by: (1) making a spectrogram; (2) normalizing the spectrogram in several ways to remove some background noises and interfering sounds; and (3) detecting sounds in the 100–1000 Hz range with a minimum duration of 1 s. The second stage, slow but accurate, operated on the sounds extracted by the first stage and classified them as being either harbor seal roars or not. Classification was done by measuring a variety of acoustic characteristics—duration, frequency span, amplitude variation, etc.—in several frequency bands, and applying statistical pattern recognition techniques to the resulting feature vectors. Training data consisted of 1011 roar examples and 850 nonroar sounds. Recognition accuracy greater than 95% was achieved, with the principal errors occurring because of close resemblance between seal roars and breakingwaves. These results show that acoustic monitoring combined with automatic recognition can be a viable method for continuous monitoring of populations of wild animals. a)Currently at PMEL, 2115 S. E. OSU Dr., Newport, OR 97365.