Call and component evaluation for improved performance of recognition of killer whale individuals

The objective of this experiment was to determine the contribution of the initial broad band component of the SD1(1a/b) vocalization towards recognition of individual killer whales (Orcinus orca). Prior research showed classification using the SD1(1a/b) vocalization performed 23% better compared to classification using the SD3(1) vocalization. One possible theory for this observation was the presence of a broad band buzz at the initiation of the SD1 call. It was theorized the broad band buzz of the vocalization was more continuously sampling the frequency response of the vocal production mechanism, (classically described as the filter in the source-filter model of speech) and potentially contributed to the observed increase in recognition. Experiments were performed with vocalizations provided by Hubbs-SeaWorld Research Institute and consisted of 20 SD1(1a/b) vocalizations for each of four whales (2 male, 2 female). The broadband component was hand segmented from the vocalization. Classification was performed on the full and segmented vocalizations with a Gaussian mixture model, using mel-frequency cepstral coefficient feature vectors. Using the full vocalization, overall accuracy was 75 +/- 2% using a 95% confidence interval. Using only the segmented broad band component, overall accuracy was 56 +/- 2% using a 95% confidence interval. Chance performance was 25%. These results cannot definitively support or reject a source filter model, but do point to the need for focused research to develop appropriate feature vectors for individual identification using acoustic cues.