Computer classification of dolphin whistles requires a number of steps from detection of a whistle in background noise to classifying the whistle among a database of whistle types. Detection was achieved by using a broadband noise reduction technique and a filter for tracking relatively slowly changing FM tones. After whistle detection, its characteristic time-frequency-intensity contour was extracted using a tracing algorithm with an 'inertial' following rule to avoid crossover when multiple whistles were present. The data requirements for each extracted contour were reduced using an encoding technique that splits each whistle into segments based on frequency slopes and using a curve fitting routine to represent the contour within each segment. Classification is achieved using hidden Markov models to represent the possible segment sequences within each class, and three distance measures based on the segment curves. In this way class similarity percentages can be calculated for each class, and a candidate whistle can be assigned to the most probable class.