Although high-quality synthesis-by-rule systems produce consonants which are easily and readily distinguishable in controlled tests, these same segments often sound unnatural in running speech. Many have attributed this lack of segmental naturalness to rules which produce "overcued" or "overarticulated" consonants. In human speech, however, it has been found that consonantal differences are often cued by the quality of adjacent vowel allophones. This paper will discuss how both the naturalness and intelligibility of consonants can be enhanced by incorporating a large number of vowel allophones into a text-to-speech system.