The problem of recognizing phonological variations in the speech input has triggered numerous treatments in speech processing models. Two areas of current controversy are the possibility of phonological underspecification in the mental lexicon and the nature of the mapping mechanism from the speech signal to the abstract lexical entry. We present data from cross-modal repetition priming experiments in English designed to test the differing predictions of speech recognition models regarding tolerance of phonological mismatch in speech. The results show effects of underspecification, supporting models using underspecified lexical entries (e.g. Lahiri & Marslen-Wilson, 1991; Reetz & Lahiri, 2001) over those using fully specified lexical entries (e.g., McClelland & Elman, 1986). The results also show no effects of context, thus favoring context-independent mapping, such as that in the three-way matching mechanism of the Lahiri and Reetz FUL model (Lahiri, 1999; Reetz, 1998, 1999; Reetz & Lahiri, 2001) over context-dependent mapping used in phonological inference rules (Gaskell & Marslen-Wilson, 1996, 1998).