Bottlenose dolprlin (Tursiops trunratus) whistles are currently studied by subjective visual comparison of whistle spectrograms. This thesis describes the novel use of stochastic modeling to automate the comparison of dolphin whistles and to yield an objective, quantitative measure of whistle similarity. The relationship of bottlenose dolphin whistle production to a model of human speech production is discussed, providing a basis for the use of human speech recognition techniques for creating whistle models.

Discrete hidden Markov models based on vector quantization of linear prediction coefficients are used to create whistle models based on statistical information derived from a sample set of dolphin whistles. Whistle model comparison results are presented indicating that evaluation of bottlenose dolphin whistles via hidden Markov modeling provides an objective measure of similarity between whistles. The results also demonstrate that hidden Markov models provide robustness against the effects of temporal and frequency variance in the comparison of whistles. The extensibility of stochastic modeling techniques to other animal vocalizations is discussed and possibilities for further work in areas such as the determination of possible structural components, similar to phonemes in human speech, is provided.