摘要
矢量量化(VQ)技术在说话人识别方面得到广泛的应用。VQ码本的产生通常采用LBG算法,LBG算法不可忽视的问题之一是空包腔的处理,它对码本的质量和算法的效率都产生影响。为此提出一种优化的空包腔处理方法:对得分最大的包腔的分解是基于聚类机会均等和码字最有代表性原则下进行的,避免了再一次产生空包腔的可能性。同时该方法还是基于不损失矢量的原则,对保持矢量空间分布的完备性具有意义。该方法得到了实验的验证。
VQ(Vector Quantization)technique is widely used in the speaker recognition field.The VQ codebook is generated by LBG algorithm.In the algorithm,how to process empty voronoi cells is a key issue for its being related to the quantity of the codebook and the algorithm's efficiency.To solve the problem,an optimized process-ing method is presented in the paper.The given maximum score voronoi cell is divided based on being well able to cluster to form sub-voronoi cells and generating most suitable codewords,so that the possibility of generating empty voronoi once again can be eliminated.The method can avoid abandoning empty voronoi cells to keep the o-riginal vectors' integrity of space distribution.The method is testified by the experiments.[
出处
《电声技术》
北大核心
2004年第3期42-44,共3页
Audio Engineering