This paper describes a new method of adapting a HMM speaker independent recogniser trained on telephone quality speech data to make it robust to background speech-like noise. The method is based on the parallel combination of a speech enhancement scheme based on masking properties of the auditory system and the HMM adaptation technique recently described by Nolazco Flores and Young [1]. The proposed enhancement algorithm uses a variation of the generalized spectral subtraction method and incorporates in it a criterion based on the human perception. Previous work [2] has demonstrated that this algorithm succeeds in finding the best trade-off between noise reduction and speech distortion in a perceptual sense. In this paper we show that this method improves also noise compensation techniques in which the parameters of corresponding pairs of speech and noise states are combined to yield a set of compensated parameters. The originality of the proposed algorithm resides in the fact that instead of using fixed parameters for the noise compensation, the noise masking threshold is used to control the enhancement and recognition processes adaptively, frame-by-frame, hence helping to find the best trade-off. The paper describes an evaluation of the presented scheme using the "Polyphone Swiss Romand" and Noisex-92 databases.