A fractional bit allocation algorithm based on Mixed Excitation Linear Prediction

In order to improve the performance of vector quantization based on different patterns, a fractional bit allocation algorithm for unvoiced/voiced(U/V) and pitch parameters based on Mixed Excitation Linear Prediction (MELP) has been proposed in this paper. This algorithm first quantizes U/V with integral bit allocation as the pattern index of pitch and gets every quantization distortion of pitch under each pattern with the allowable codebook size, then utilizes dynamic programming to get the optimal codebook size allocation under a given quota of bit number and calculates the corresponding total quantization distortion. The same procedure is conducted again with the number of patterns decreasing by one until the distortion increase rate of U/V exceeding the distortion decrease rate of pitch. Simulation results show that the method proposed here can reduce at least 25% of the quantization distortion of pitch and increase Mean Opinion Score (MOS) of the synthesized speech by 0.03.