Performances of an automatic speech recognition system degrade when test and training conditions do not match. Classical Stochastic Matching (SM) method proposes an off-line estimation of a compensation function that maximizes the likelihood of the compensated speech, given the optimal sequence of models proposed by the recognition process. We developed a new frame-synchronous technic based on SM : compensation is performed in parallel with the recognition. This is suitable to cope with slowly varying noise. We proposed two additional versions of our approach: -a tree structure of transformations is used to build a state-dependant non-linear compensation function. This is motivated by the fact that similar observations will be affected similarly by the environment. -a surveillance process monitoring the fluctuations in the environment is used to trigger the reinitialisation of the compensation process. This enables our algorithm to cope with environments experiencing sudden occurrences of noise.