论文部分内容阅读
在矢量量化孤立字识别系统中,对于识别字典中每一个单字,都要分别建立与这些单字对应的码本。本文提出一种按音节递归训练码本的算法,它的计算量是训练语音帧数的线性函数。而传统的LBG聚类算法的运算量则随训练矢量增加呈指数率增长。本文比较了这两种算法的失真特性,得出结论:新方法需要n+1次发音训练,就能达到全搜寻算法n次发音训练的失真性能。最后,本文给出矢量量化孤立字识别器的实验结果。
In a vector quantization isolated word recognition system, for each word in the recognition dictionary, codebooks corresponding to the words are respectively established. This paper presents a recursive training codebook algorithm based on syllables, the calculation of which is a linear function of the number of training speech frames. The traditional LBG clustering algorithm with the training vector increases exponentially. This paper compares the distortion characteristics of the two algorithms and concludes that the new method requires n + 1 pronunciation training to achieve the distortion performance of n-th pronunciation training of the full search algorithm. Finally, the experimental results of vector quantization isolated word recognizer are given in this paper.