A new method for speech parametric coding at medium output rate is presented. The availability of new electronic components changes the hypotheses assumed in previous studies. A parametric coding scheme with medium output rate now seems appropriate for rule-based synthesis. The weak spot in most parametric synthesizers lies in the poor model used to represent the excitation signal. Our approach consists in coding, in the frequency domain, the residual signal of a linear prediction analysis. The residual signal is analysed as a function of frequency through a sub-band splitting. Particularly, voicing is analysed in each band and the decision is progressive from voiced to unvoiced. This parametric, frame based, representation of the residual signal, associated with the linear prediction coefficients, is used for transmission or storage purposes. The synthesizer builds the synthetic residue by generating spectra from these parameters. These spectra are combined through the overlapp-add algorithm to obtain a time domain signal. Our system thus allows complex filtering and phase control and provides a robust representation and a good quality while preserving the flexibility inherent to parametric coding schemes.