论文部分内容阅读
为了在大词汇量连续语音识别(LVCSR)系统中能够利用段长信息,该文按树状组织发音词典,利用语言模型预测技术,基于最大似然状态序列(M LSS)算法,给出了采用基于段长分布的隐含M arkov模型(DDBHMM)的LVCSR系统的二元文法语言模型的单步搜索算法。实验结果表明,尽管单步搜索的替代错误率高于双步搜索,但单步搜索的插入和删除错误率都比双步搜索要低,总体性能上单步搜索要好于双步搜索。同时,DDBHMM能较准确地利用了语音信号中的状态段长信息,采用DDBHMM的LVCSR系统比采用经典的齐次HMM的系统有更好的识别性能。
In order to make use of the segment length information in large vocabulary continuous speech recognition (LVCSR) system, the paper presents a tree-based pronunciation dictionary using the language model prediction technique and the Maximum Likelihood Sequence (MLSS) algorithm, A Single Step Search Algorithm for Binary Grammar Language Model of LVCSR System Based on Sectional Length Distribution of Implied M arkov Model (DDBHMM). Experimental results show that although the replacement error rate of single-step search is higher than that of double-step search, the single-step search has lower insertion and deletion error rates than double-step search, and the overall performance of single-step search is better than double-step search. At the same time, DDBHMM can make more accurate use of the state length information in voice signals. The DDCSH system using DDBHMM has better performance than the system using classical homogeneous HMM.