矢量量化孤立字识别的研究

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。

A Vector Quantization Based Isolated Word Recognizer

In vector quantization(VQ)approach to discrete utterance recognition,a set of codebooks are generated,one for each word in the recognition vocabulary.To achieve more reliable and robust eodebooks for discrimination,the codebook is often designed from a training sequence containing several repetitions of the same word utterances.This paper introduces a recursive clustering algorithm for efficient codebook generation,which updates the codebook from an utterance to an utterance. Experiments show that the distortion performance of the recursive procedure is very close to that of the full search procedure.In other words,the former requires roug- hly one more training utterance than the latter to achieve the same level distor- tion for open.test.But the computation efforts have been substantially reduced. Finally,several recognition results of the VQ based recognizer are also presented.