Birdsong shares many features with spoken language while remaining free from coarticulatory effects. Birdsong syllables are stereotyped, acoustically complex, and follow complex syntactic rules. These aspects make birdsongs useful test beds for understanding how sparse codes can support recognition of vocal units in auditory scenes. In this talk, we introduce a recognition system that uses ultra-sparse sequences of spikes to represent birdsongs. We find that it recognizes syllables at low error rates while using only a few seconds worth of syllables for training, a 10-fold reduction in training set duration from the state-of-the-art. Overall, our results emphasize the important role of neuroscience in guiding research in machine intelligence.