A bottlenosed dolphin (Tursiops truncatus) was trained to mimic computer-generated "model" sounds, using a whistle mode of vocalization. Prior to training, the whistle sounds of this dolphin were limited to a few stereotyped forms, none of which resembled the model sounds. After training, high-fidelity imitations were obtained of model sounds having (a) moderately or widely swept, slow-rate frequency modulation (1-2 Hz), (b) narrowly or moderately swept frequency modulation at moderate to rapid rates (3-11 Hz), (c) square-wave frequency transitions, and (d) unmodulated (pure-tone) waveforms. New models, not heard previously, could be mimicked immediately, often with good fidelity, including mimicry of amplitude variation that had not been explicitly reinforced during training. Subsets of familiar models were mimicked with high reliability in repeated tests. In additional training, control of the mimic response was transferred from the acoustic model to objects shown the dolphin (e.g., a ball or a hoop) so that, in effect, the dolphin gave unique vocal labels to those objects. In a test of accuracy and reliability of labeling, correct vocal labels were given on 91% of 167 trials comprised of five different objects presented in random order. The dolphin's ability for vocal mimicry compared favorably with that of the more versatile mimic birds, and it contrasted sharply with the apparent lack of vocal mimicry ability in terrestrial mammals other than humans. The ability to label objects vocally was similar to abilities shown for some birds and similar, in principle, to abilities of great apes trained in visual languages to label objects through gestures or other visual symbols.