| 摘要: |
| 提出了一种具有强分类能力的离散HMM(hiddenMarkovmodels)训练算法.该算法利用矢量量化技术将来自不同话者的训练数据进行混合训练,以生成包含各个话者特征的话者特征图案.用该特征图案代替经典的离散HMM中的VQ码本,可以提高观察值符号序列的模式辨识能力,从而提高了离散HMM的分类能力.给出了该方法用于文本有关的话者识别的实验结果,表明该算法可提高系统的识别性能,并要降低HMM对训练集大小的依赖程度,且识别时计算量明显小于经典HMM训练算法,具有较大的实用价值. |
| 关键词: 离散HMM(hidden Markov models) 分类能力 特征图案 矢量量化 鲁棒性 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金资助项目(69872036) |
|
| An Algorithm with Strong Classifying Ability for Discrete HMM Training |
|
FANG Shao wu,DAI Bei qian,LI Xiao han
|
| Abstract: |
| A discrete-HMM training algorithm which has strong ability of pattern classification is presented in this paper. By VQ (vector quantization) technique, this algorithm trains data from all speakers in mixed mode to generate the speaker characteristic pattern, which includes features of all speakers. By substituting the VQ code\|book in conventional discrete-HMM with characteristic pattern, the ability of pattern classification for observation symbol sequence is enhanced, therefore the classifying ability of discrete-HMM is improved. The experimental results show that the algorithm can improve the system's recognition performance, and reduce the dependence extent of HMM on the scale of training set. Moreover, the calculation quantum of this algorithm in recognition stage is obviously less than that of conventional HMM training algorithm, therefore it has higher practical value. |
| Key words: discrete hidden Markov model classifying ability characteristic pattern vector quantization robustness |
