This paper describes a new approach for modelling allophones in a speech recognition system based on hidden Markov models. This approach allows to model in detail, and with a limited amount of parameters, the different acoustical realizations of the sounds by integrating left and right context-dependent transitions as well as acoustical targets in the basic units. Phonetic knowledge is used to define the structure of the models, and a standard training procedure determines the optimum value of the parameters. The efficiency of the approach is demonstrated both in a multispeaker mode, with a 500-word vocabulary, and in a speaker-independent mode with several other data bases recorded through the telephone network by more than 500 speakers. The improvement resulting from the use of temporal derivatives are also compared for several kinds of basic units. Keywords: Speech recognition, Markov modelling, Modelization of allophones, Basic units, Temporal derivatives.