In this report, we develop a method, called HMM composition to cope with the problem of speech recognition in a noisy environment avoiding the tedious training of noisy HMMs. We then consider its application to a speech recognition system based on LPC cepstrum parameters. The method was tested against a variety of noises, stationary and non-stationary with signal to noise ratios ranging from 0dB to 20dB and provides an error reduction over 75% comparing with the clean-speech HMM. It is believed that this technique, by its efficiency, its flexibility and its adaptability to new noises and SNRs could constitute the heart of a real-time speech recognizer robust to noise.