一种基于模糊聚类的矢量量化码书生成算法

A Novel Algorithm of Vector Quantization Codebook Design Using Fuzzy Clustering Theory

码书设计是矢量量化中的关键技术.为此,针对经典LBG算法对初始码书敏感的缺陷提出一种基于模糊聚类的码书生成算法.为了提高收敛速度,首先设定距离门限的初始值,然后依次循环逐级递减调整以减少迭代次数.逐级调整门限的也可以降低新聚类生成的速度,从而得到更好的更具有典型性的码书;此外,通过对胞腔中矢量按从大到小的顺序择优选取,设计出的码书性能更好,更加接近全局最优.将该算法应用于移动语音频编码标准中线谱频率矢量量化的码书训练,与LBG算法的对比实验结果表明,该算法在主客观质量评价方面都有效地提高了语音频编码算法的性能.

The design of the codebook is of great importance in the vector quantization. This paper, pointing to the drawbacks of the classic algorithm LBG, proposes a novel efficient algorithm of designing codebook based on the fuzzy clustering theory. Through initialization of distance threshold and the following adaptation of reducing threshold for each level, the iteration time is reduced and the convergent speed is then improved. At the same time, the adapta- tion of reducing threshold for each level can induce the generation of new clusters, thus resulting in more typical codehook vector. At last, through optimization selection of vectors in the decreasing order, the designed codebook will achieve better performance and be closer to global optimization. The novel algorithm is applied to LSF vector quantization codebook generation in mobile speech/audio coding, and the experiment result shows that the algorithm can improve the objective and subjective performance of the audio codec.