The warble song of male budgerigars (Melopsittacus undulatus) is an extraordinarily complex, multi-syllabic, learned vocalization that is produced continuously in streams lasting from a few seconds to a few minutes without obvious repetition of particular patterns. As a follow-up of the warble analysis of Farabaugh et al. (1992), an automatic categorization program based on neural networks was developed and used to efficiently and reliably classify more than 25,000 warble elements from 4 budgerigars. The relative proportion of the resultant seven basic acoustic groups and one compound group is similar across individuals. Budgerigars showed higher discriminability of warble elements drawn from different acoustic categories and lower discriminability of warble elements drawn from the same category psychophysically, suggesting that they form seven perceptual categories corresponding to those established acoustically. Budgerigars also perceive individual voice characteristics in addition to the acoustic measures delineating categories. Acoustic analyses of long sequences of natural warble revealed that the elements were not randomly arranged and that warble has at least a 5th-order Markovian structure. Perceptual experiments provided convergent evidence that budgerigars are able to master a novel sequence between 4 and 7 elements in length. Through gradual training with chunking (about 5 elements), birds are able to master sequences up to 50 elements. The ability of budgerigars to detect inserted targets taken in a long, running background of natural warble sequences appears to be species-specific and related to the acoustic structure of warble sounds.