论文部分内容阅读
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．
This thesis has developed a voice-activated voice command exchange system for telephone subscriber exchange. The system can automatically recognize small and medium vocabularies (<1000 words) independently of the specific person. In the research, the commonly used command sentences of the user exchange system are counted to generate the corresponding recognition grammar network. The training of recognition system adopts the compound model made up of sub-word model to strengthen the training, and uses token passing to improve the Viterbi algorithm to improve the recognition performance of the system. The thesis compares the different voice features parameters and implicit Markov model state number on the telephone speech recognition accuracy. In the research, we also develop a recognition system rejection algorithm, which commands a correct recognition rate of over 88% without rejection, and achieves over 95% recognition rate by adding a rejection algorithm at a recognition rate of 20%.