在G.729建议的共轭结构代数码激励线性预测编码(CS-ACELP)中,固定码本搜索在整个语音编码算法中占有较大比重,直接影响编码算法复杂度。全搜索算法准确度很高,但搜索量过大,而传统的脉冲序列替换搜索,搜索次数减少,但合成语音质量较差。为解决该问题,提出一种基于脉冲序列替换的改进码本搜索算法。设定循环阈值门限,对脉冲序列重置后的部分脉冲组合进行全搜索,引入双脉冲位置替换,有效地减少了搜索次数,同时提高了搜索准确度。实验结果证明,该算法在增加算法复杂度较少的情况下,合成语音质量有明显的改进。
The fixed-codebook searching complexity occupies larger proportion of the whole encoding complexity in G.729,which influences the efficiency of the encoding algorithm.The nested-loop search can obtain the optimal codebook vector,but it requires large computational load.Although the traditional pulse replacement search method can reduce the search times greatly,the synthesized speech quality degrades.In order to solve this problem,gave a new approach based on pulse replacement method.In each track of the reordered pulse sequence,searched the codebook partly in accordance with a fixed threshold which related to the loop times.Introduced a double pulse replacement method,which could increase the accuracy of the codebook search with low search load.Tests show that the synthesized speech quality has been greatly improved with the proposed algorithm.Meanwhile,the computational complexity increases slightly compared to the traditional pulse replacement method.