The categorization of word-final phonemes provides a test to distinguish between an interactive and an autonomous model of speech recognition. Word-final lexical effects ought to be stronger than word-initial lexical effects, and the models make different reaction time (RT) predictions only for word-final decisions. A first experiment found no lexical shifts between the categorization functions of word-final fricatives in pairs such as fish-fiss and kish-kiss. In a second experiment, with stimuli degraded by low-pass filtering, reliable lexical shifts did emerge. Both models need revision to account for this stimulus-quality effect. Stimulus quality rather than stimulus ambiguity per se determines the extent of lexical involvement in phonetic categorization. Furthermore, the lexical shifts were limited to fast RT ranges, contrary to the interactive model's predictions. These data therefore favor an autonomous bottom-up model of speech recognition.