A recognition system must include structures which account for the different aspects of possible phonological variability in the incoming speech signal. To aid in dealing with variability at the phonological level, a grapheme-to-several-phoneme strings module, VARION0, has been developed at LIMSI. Tests of this system showed the necessity of creating a hierarchical structure to order phonological variations, according to a fixed set of criteria, eliminating infrequent phenomena, thus bringing the number of possible strings down to a reasonable size. The resulting new module, VARION1, was tested on transciptions of recorded speech (BDSONS-30 speakers and a variable speech rate database).