n this paper, we deal with a new method to globally optimize a Multi-Band Speech Recognition (MBSR) system. We have tested our algorithm with the TIMIT database and obtained a significant improvement in the accuracy over a basic HMM system for clean speech. The goal of this work is not to prove the effectiveness of MBSR, what has yet been done, but to improve the training scheme by introducing a global optimization procedure. A consequence of this method is that the models are no longer phone-models, but define new classes in the phonetic space which might better model the acoustic information carried by each band.