基于DDBHMM的LVCSR系统的单步搜索算法

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。

One-stage search algorithm for large vocabulary continuous speech recognition based on DDBHMM

In order to use duration information in a large vocabulary continuous speech recognition(LVCSR) system,the pronunciation dictionary is organized as a tree and the language model look-ahead technique is adopted.Based on the maximum likelihood states sequence algorithm,the one-stage search algorithm for the LVCSR using the duration distribution-based hidden Markov model(DDBHMM) in proposed when the Bigram language model is used.Tests show that,although the two-stage search algorithm has a lower substitute error rate than the single stage one,the insertion errors and deletion errors are both higher than that of the single-stage search.The one-stage search algorithm is,therefore,better than the two-stage search in terms of overall performance.Since the DDBHMM accurately describes the state duration of the speech signals,the DDBHMM system has better performance than system using homogeneous HMM.