Abstract: Linear predictive technique has been widely used for speech signal processing, especially for low bit rate vocoder in recent years. However the conventional fixed frame processing does not suit well with speech signal for its nonstationary behavior. Also it is not reliable to exltraot pitch from the parameters （e.g. residual） obtained by this method. Due to these reasons, speech quality can be adversely affected. In order to improve the quality and intelligibility, in this paper it presents improvement to the following three aspects: （1） Using adaptive ladder form algorithm instead of conventional LPC to obtain more accurate estimation of the vocal-tract transfer function （2） Adapting non-uniform sampling of speech parameters with constrained interval to improve performance of synthesized speech during fast transients. （3） Improving Gold-Rabiner pitch extraction technique for saving time and reliable. The computer simulated results show that the 2.4 kb/s vocoder gives the synthesized speech with good intelligibility and quality, and the quality of the 1.2 kb/s one is not degraded seriously. Transmission over fixed rate channels can be accomplished, using transmit and receive buffers.