In this contribution we propose an improvement of the multiband excitation vococoder on the basis of preserving temporal and spectral features of transients. Proper representation of unvoiced-voiced transitions is achieved by means of reliable detection of the point where harmonics are born and application of specific analysis and synthesis procedures for the frames placed before and after this point. Description of temporal features requires the definition of new parameters, their extraction and quantization having been studied. Additional methods for parameter reduction and quantization will also be addressed: