A lexical access using a two-stage algorithm based on the description of the lexicon into two types of units is proposed. To access to a sub-dictionnary, phonetic classes are used to build an HMM. Then this sub-dictionary is modeled by a temporary HMM, with pseudo-diphones units, to produce the recognized word. An acoustic pre-processing adapted to the pseudo-diphone decomposition and an adequate "HMM Compiler" toolkit are used in this application. Experiments are made to evaluate this approach with the 500 most frequents words in french language.