Speech learning involves discovering linguistically-relevant categories embedded in continuous speech. But, learninghas been investigated mostly across isolated sound tokens. Here, we investigated incidental learning across continuousmulti-talker Mandarin speech in the context of a videogame in which participants behavior was directed at navigating avirtual environment, not speech learning. Unbeknownst to the native-English participants, acoustically-variable Mandarinkeywords were embedded in the continuous sentences, and were associated with game actions and events. Participantswere not informed about the keywords, made no categorization decisions, and received no overt feedback. Post-trainingresults indicated robust keyword learning that persisted at least 10 days. Further, the electrophysiological N100 componentevoked by keywords during passive listening to continuous Mandarin was greater post-training than pre-training. This neu-ral enhancement was not observed for equally-frequent control keywords unassociated with game behaviors. Participantslearned functionally-relevant non-native speech categories incidentally from continuous speech input across considerableacoustic variability.