While it is generally assumed that human speech perception starts with the identification of the smallest units, i.e., phones, followed by lexical access, the great variability found in the acoustic characteristics of continuous speech on the one hand and the apparent ease of human listeners in coping with the variability on the other call for re-examination of the conventional view. This paper describes a few experiments conducted to examine the influence of context and knowledge on the units of recognition as well as that of familiarity on the ease of lexical access. A tentative model is then presented for the human processes of spoken language perception.