Abstract: The offset of gain filter is dealt unsuitably with a constant 32dB in G.728 standard. That is, the short-term stationary speech signal during the duration 20~30ms is unsuitably supposed as stationary at whole time axis. It is necessary to design optimizing hybrid window for improving gain filter. In this paper, the exact expression of excitation gain is derived and an estimation method of SNRav used to optimize the parameters of hybrid window is presented. On the basis of above, this paper combines the Jayant adaptive quantization and G.728s gain prediction to propose a scheme fit for designing gain codebook of Low Delay Code Excited Linear Prediction(LD-CELP), in which the exact gain value is adaptively predicted and quantized. In this method, for adaptive quantization gain of 3 bit and 4 bit, the SNR increases by 0.5 and 6 dB separately than that of G.728. If adopting 4 bit adaptive quantization and the shape codebook of 64 codevectors, the G.728 complexity can reduce by 20% while SNR has 1 dB improvement.