Previous studies of word learning have presented the items to listeners under ideal conditions. Here we ask how listeners learn new vocabulary items under adverse listening conditions. Would listeners form acoustically-specific representations that incorporated the noise, base their representations on noise-free language knowledge, or both? To address these questions, listeners learned 16 words as labels for unfamiliar shapes presented on a computer display. During the learning phase, word-shape pairings were presented with either clear or white-noise-embedded tokens. For each word (e.g. "dabo"), another word shared consonants (e.g. "dubei") and a third shared vowels (e.g. "gapo"). Learning was assessed in a 4AFC picture-selection task. The highest accuracy and speed were achieved by listeners who experienced the same noise level at exposure and test (both clear or both noisy), suggesting that listeners' representations of noisy words were faithful to the spectral context experienced during the learning phase. Implications for word learning and recognition across a variety of listening conditions are discussed. (Contains 5 figures and 2 footnotes.)