Previous studies on starling song had shown that different categories of songs are produced and discriminated. In the present study, both playback experiments in the field and electrophysiological recordings of single cells in field L (main central auditory area) were used in order to understand the perceptual bases of category recognition. In both cases, synthesized variants were used where different parameters could be changed. A method of backward correlation was used in order to visualize and determine the precise elements in song towards which the neurons responded. A high proportion of neurons was found that responded selectively to particular features in the species-specific songs. Series of neurons with similar or complementary selectivities were found. We did not find any 'category-detector' as such but key elements appeared both in neuronal selectivity and in behavioural discriminations. The data favour the propositions of the pattern recognition theory, with a multiple feature model. (C) 2000 Elsevier science B.V. All rights reserved.