Infants develop phonetic categories by simply being exposed to adult speech. It remains unclear, however, how they handle the extensive variability inherent to speech, and how they process multiple linguistic functions that share the same acoustic parameters. Across four neural network simulations of lexical tone acquisition, self-organizing maps were trained with continuous speech input of increasing variability. Robust tonal categorization was achieved by tracking the velocity profiles of fundamental frequency contours. This result suggests that continuous speech signal carries sufficient categorical information that can be directly processed, and that dynamic acoustic information can be used for resolving the variability problem.