几种小训练样本集的数字语音识别模型的比较性研究  

The Comparative Research of Some Digit Speech Recognition Models Based on Small Training Corpus

作　　者：贺苏宁[1] 虞厥邦[1] 

机构地区：[1]电子科技大学电子工程学院,成都610054

出　　处：《计算机科学》2005年第9期170-175,共6页Computer Science

摘　　要：本文通过对小训练样本集的基于DTW结构的数字语音识别模型的比较性分析,指出其存在的三个一般性问题:(1)DTW逐帧匹配模式割裂了观测向量序列的内在联系;(2)压扩观测向量序列造成局部信息使用的不均匀;(3)计算复杂度高,识别率低。为了解决这些问题,我们提出了基于数字语音时频信息整体结构的单特征向量识别模型。这种模型完整地利用了观测向量序列的全部信息,结合置信度评估和自适应反馈学习之后可及时地吸收测试向量携带的新的环境特征信息,调整识别模型结构。该模型的错识率较之最好的基于DTW结构的混合城模型的错识率降低50％以上,计算复杂度则是固定帧长模型的 13.12％。By analyzing comparatively digit speech recognition models based on DTW structure with small training corpus, we point out their three general problems..（1） Such a DTW matching mode of one by one frame splits the internal relation within observation vector sequence. （2） Compressing or expanding observation vector sequence makes local information being used asymmetrically. （3） High complex computation and low recognition rate. In order to solve these problems, we proposed a single feature vector recognition model based on whole time-frequency information structure of digit speech. The model makes full use of all the information of observation vector sequence. It can absorb in time the new environmental features being taken by testing vector and adjust its structure by combining the confidence measure with self - adaptive feedback learning. The error recognition rate is less than half of that the best mixed model＇s with DTW structure. It takes only 13. 12 percent of time to fulfill the computation compared with that of the fixed - length frame model.

关 键 词：训练样本集 数字语音识别模型 置信度评估 自适应反馈学习 DTW 匹配模式 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]