摘要
本论文针对电话用户交换机研制了一个声控语音命令交换系统．该系统能够实现与特定人无关中小词汇量（＜１０００词）连续命令语音自动识别．研究中统计了用户交换系统的常用命令语句，生成相应识别文法网络．识别系统的训练采用由子词模型构成的复合模型进行强化训练，识别采用令牌传递式改进Ｖｉｔｅｒｂｉ算法，提高系统的识别性能．论文比较了不同语音特征参数以及隐含马尔可夫模型状态数对电话语音识别精度的影响．研究中还开发识别系统拒识算法，在无拒识情况下命令正确识别率达到８８％以上，通过加入拒识算法，在２０％的据识率情况下其识别率可达９５％以上．
In the paper the speech recognition system is developed based on private automatic branch exchange(PABX).This system can recognize speaker independent continue speech commands over telephone channel based on medium and small vocabulary(<1000 words).Statistics of the different sentence pattern of speech commands has been done and the corresponding recognition networks has been established.The system uses enhancive training combined with a embedded composite model which concatenate the sub word models,and improved token passing Viterbi algorithm,which improve the recognition accurate of the system.In the paper recognition accurate are compared according to different speech characteristic parameters and the state numbers of Hidden Markov Model.The rejecting recognition algorithm are studied.The recognition accurate is over 88% without rejection and over 95% with 20%rejection of utterances.
出处
《电子学报》
EI
CAS
CSCD
北大核心
1999年第1期5-7,共3页
Acta Electronica Sinica
基金
国家自然科学基金
国家863项目
