Recent work demonstrates that talker characteristics can be used as predictive cues for spoken word recognition. However, abstractionist accounts suggest that talker information is usually stripped away or normalized based on preceding speech material. A contrasting account is that listeners never normalize, instead storing detailed, acoustically-varied instances of words. These varied instances then facilitate recognition of words in a vast variety of voices and accents. We present data suggesting that such irrelevant acoustic characteristics of word forms (talker-varying acoustic attributes) are not normalized, but are instead encoded, when learned in fluent speech context. Experiment 1 replicates recent demonstrations of talker specificity in word recognition. Using the same set of words in carrier sentences, Experiment 2 finds that learners still encode talker information even though in principle they could easily normalize away talker-based acoustic variability.