This paper presents a set of solutions aiming to improve the performances of speech recognition systems used in adverse environments, especially in cellular phone noise conditions. Hidden Markov Models achieve very good performances in speaker independant small vocabularies speech recognition, and offer realistic solutions for creating interactive vocal servers, but the lack of robustness to environment changes is their principal weakness. Many authors have proposed either robust recognition methods, or noise cancellation systems, but none of these methods achieves good results for cellular phone specific noise, which is often drastically corrupted by ambient noise, transmission noise and impulsive noise. The novel solution we present here consists in a specific HMM modelling of multi-noise contexts, training the system in various noise environments, each of them supported by a component of the system, either a subsystem, or a subset of transitions and density laws.