多级矢量量化中的码本共享

为了解决语音参数编码算法中多级矢量量化中码本尺寸过大,存储量过大,导致搜索复杂度大的问题,提出了多级矢量量化中的码本共享的迭代算法。该算法基于多级矢量量化中各级待量化码矢之间的相似性,采用模拟退火算法,通过迭代得出共享变换系数。在1.2 kb/s的正弦激励线性预测声码器中,采用该算法对线谱对参数进行多级矢量量化。测试结果表明:在共享级别选择恰当时,可降低存储量20%,同时重建语音谱失真损失约为0.02 dB,可见该算法可以有效降低码本容量,同时对语音质量影响极小。

Codebook sharing in multi-stage vector quantization

An iterative codebook sharing algorithm for MSVQ(multiple stage vector quantization) was developed to reduce the codebook size for storage and searching.The algorithm uses the simulating annealing method to calculate the codebook sharing transform coefficients,based on the shape similarity between different stages of the codebook.Test results in a 1.2 kb/s SELP(sinuous excitation linear prediction) vocoder which uses codebook sharing in the MSVQ of the LSP(line spectrum pairs) parameter show that if codebook sharing level is properly selected,the codebook size is reduced by 20% with a spectral distortion loss of only 0.02 dB which is neglectable.Thus,the algorithm can efficiently reduce the codebook size with neglectable distortion of the speech quality.