基于连续语音识别的码本数据信息优化研究

Optimization of Codebook Data Information Based on Continuous Speech Recognition

在语音识别技术中,矢量量化起着非常关键的作用。矢量量化的码本设计过程中,经典的LBG算法对初始码本非常敏感,初始码本的选取不但影响迭代的收敛速度,很大程度上还影响最终码本的性能。本文提出一种新的初始码本生成算法,通过距离调节参数进行迭代来更细致地划分矢量空间,进而改善码本质量。新算法与其他算法进行对比,在类间离散度、矢量量化不均匀度和收敛性等性能指标上,都具有明显优势,同时将新算法得到的初始码本经LBG算法再次优化后,应用于非特定人汉语连续语音识别,实验结果证实了新算法的有效性。

Vector Quantization plays a very important role in speech recognition.In the process of codebook design of vector quantization,the classical LBG algorithm is very sensitive to the initial codebook.The selection of the initial codebook not only affects the convergence speed of iteration,but also greatly affects the performance of the final generated codebook.In this paper,a new initial codebook generation algorithm is proposed,which partitions the vector space more carefully by iterating the distance adjustment parameters,and then improves the quality of codebook.Compared with other algorithms,the new algorithm has obvious advantages in such performance indexes as inter-class dispersion,vector quantization non-uniformity and convergence.At the same time,the initial codebook obtained by the new algorithm is optimized again by LBG algorithm and applied to speaker-independent Chinese continuous speech recognition.The experimental results show that the new algorithm is effective.