| 训练方法对汉语数字串识别率影响的研究 |
| |
| 引用本文： | 汤霖,蔡莲红.训练方法对汉语数字串识别率影响的研究[J].湘潭大学自然科学学报,2003,25(3):16-20. |
| |
| 作者姓名： | 汤霖 蔡莲红 |
| |
| 作者单位： | 1. 江门教育学院计算机系,广东,江门,529000
2. 清华大学计算机系,北京,100084 |
| |
| 摘 要： | 在语占识别系统中，都是通过提取特征向量来计算待识语音与模型之间的概率或距离，然后根据最大概率或最小距离判断待识语爵的类别．对大量实验数据的观察发现：特征向量的各维对语音的表达能力是不一样的，同时特征向星在语音的时间轴上表达能力也不一样。根据这种特性，提出了三种训练算法：在训练中计算出加权矩阵，以此来加强易混淆数字间的本质区分特征，减弱随机特征，在汉语数字串识别实验中，得到了比较理想的实验结果，错误下降40．8％，系统识别率上升到94．08％。
|
| 关 键 词： | 语音识别 汉语数字串 洲练算法 加权矩阵 |
| 文章编号： | 1000-5900(2003)03-0016-05 |
| 修稿时间： | 2003年6月20日 |
Research work on the influence of training method in mandarin connected digit speech recognition |
| |
| TANC Lin,CAI Lian - hong.Research work on the influence of training method in mandarin connected digit speech recognition[J].Natural Science Journal of Xiangtan University,2003,25(3):16-20. |
| |
| Authors: | TANC Lin CAI Lian - hong |
| |
| Abstract: | In speech recognition system, the probability or distance between the speech to be recognized and the model is calculated by extracted feature vectors, the speech is recognized as the same kind that who has the maximum probability or minimum distance. By lots of experiments , it is observed that feature vectors is different in describing different phoneme in different characters and in different frames. By this ,this paper provide three training methods using weight matrix to emphasize the inherent difference and decline the random difference. In mandarin connected speech recognition system, the recognition rate has been increased to 94.08% by using those methods, and the recognition error rate has been decreased by 40.8% . |
| |
| Keywords: | speech recognition mandarin connected digit training method weight matrix |
| 本文献已被 CNKI 维普 万方数据 等数据库收录！ |
