In this paper, we have gathered new results obtained with our Multi-Band system on the TIMIT database corrupted by additive noise. The robustness of the Multi-Band system to several kinds of noise is studied in details. The system is robust to spectral-limited noise, as it has already been shown in previous papers, but it is also robust to broad-spectral noise as long as the full-band is added as an additionnal stream to the Multi-Band system. A more robust version of this system is then obtained when training the recombination module in white noise. Thus, robustness to other kinds of noise is increased. Finally, we present the first results obtained with a new training algorithm which globally optimizes the Multi-Band system. Regarding the robustness of the method, these preliminary results are encouraging.