Automatic speech recognition works quite well in clean conditions, and several algorithms have already been proposed to deal with stationary noise. The next challenge probably consists to compensate for non-stationary noise as well. This work studies this problem by proposing and comparing two adaptations of the Parallel Model Combination (PMC) algorithm for non-stationary noise. A third method, derived from the missing data framework, is further proposed and compared to the two previous ones. In musical noise, experimental results show an important improvement of the recognition accuracy for one PMC-derived algorithm, compared to the non adapted system. The missing-data algorithm also performs quite well, despite its simplicity and the strong assumptions he is using.