In order to recognize spoken words, listeners must map sensory information from the acoustic input onto stored lexical entries. Because the speech signal is continuous, listeners must segment the speech stream in order to recognize words. To accomplish the task of segmentation listeners use their tacit knowledge of a wide range of patterns in their native language including cues from allophonic variation, phonotactic constraints, transitional probabilities, lexical stress etc. Among those cues, there is now a growing body of evidence suggesting that fine-grained acoustic information is available for lexical access and used for segmenting the speech stream. Although it is generally agreed that acoustic cues are used on line to segment the speech signal and to bias lexical access, some important questions remained unanswered. First we ought to know whether these cues are robust enough to be used in the context of multiple productions of the same segmentation as speech is by nature variable and listeners are never exposed to invariant speech. The second important open question is that of the timing of the use of the cues.