Computational models can reflect the complexity of human behaviour by implementing multiple constraints within their architecture, and/or by taking into account the variety and richness of the environment to which the human is responding. We explore the second alternative in a model of word recognition that learns to map spoken words to visual and semantic representations of the words’ concepts. Critically, we employ a phonological representation utilising coarse-coding of the auditory stream, to mimic early stages of language development that are not dependent on individual phonemes to be isolated in the input, which may be a consequence of literacy development. The model was tested at different stages during training, and was able to simulate key behavioural features of word recognition in children: a developing effect of semantic information as a consequence of language learning, and a small but earlier effect of phonological information on word processing. We additionally tested the role of visual information in word processing, generating predictions for behavioural studies, showing that visual information could have a larger effect than semantics on children’s performance, but that again this affects recognition later in word processing than phonological information. The model also provides further predictions for performance of a mature word recognition system in the absence of fine-coding of phonology, such as in adults who have low literacy skills. The model demonstrated that such phonological effects may be reduced but are still evident even when multiple distractors from various modalities are present in the listener’s environment. The model demonstrates that complexity in word recognition can emerge from a simple associative system responding to the interactions between multiple sources of information in the language learner’s environment.