Vocal communication requires the formation of acoustic categories to enable invariant representations of sounds despite superficial variations. Humans form acoustic categories for speech phonemes, enabling the listener to recognize words independent of speakers; animals can also discriminate speech phonemes. We investigated the neural mechanisms of this process using electrophysiological recordings from the zebra finch secondary auditory area, caudomedial nidopallium (NCM), during passive exposure to human speech stimuli consisting of two naturally spoken words produced by multiple speakers. Analysis of neural distance and decoding accuracy showed improvements in neural discrimination between word categories over the course of exposure, and this improved representation transferred to the same words by novel speakers. We conclude that NCM neurons formed generalized representations of word categories independent of speaker-specific variations that became more refined over the course of passive exposure. The discovery of this dynamic encoding process in NCM suggests a general processing mechanism for forming categorical representations of complex acoustic signals that humans share with other animals.