Cochlear Implant (CI) is a device meant to recover hearing abilities for patients suffering of total bilateral cophosis. In this study we try to identify phonemes that usually have a low CI recognition rate, using mel-frequency cepstral coefficients (MFCC) technique. In order to analyze them we had to structure a special signal database: we registered phonemes passed by the CI testing device. By a technique of accentuating certain frequency bands - depending on the recognized phoneme - we intend to improve CI performances. Clustering is realized by the help of a usual two-layer MLP neural network. In parallel, we extract, using the same step as in the cochlear implant device technique already implemented, a new set of MFCC coefficients from speech, and compare them. Using fuzzy functions in computing energy on frequency bands results are slightly better according to M. Costin et al.(2002).