| 与文本无关的复合策略说话人辨识系统 |
| |
| 引用本文： | 牟晓隆,胡起秀,吴文虎.与文本无关的复合策略说话人辨识系统[J].清华大学学报(自然科学版),1997(3). |
| |
| 作者姓名： | 牟晓隆 胡起秀 吴文虎 |
| |
| 作者单位： | 清华大学计算机科学与技术系，智能技术与系统国家重点实验室 |
| |
| 基金项目： | 国家“八六三”高科技项目 |
| |
| 摘 要： | 为获得较高的说话人辨识正确率，同时减小辨识系统的时空开销，提出了一种复合策略的辨识系统。采用长时平均频谱作为粗识的特征，定义了相应的辨识判别准则。建立ｍｅｌ－倒谱特征的高斯混合模型（ＧＭＭ）进行第二步辨识。给出了ＧＭＭ求解算法的一种简便推导，着重研究了判别阈值，预加重系数，ＧＭＭ阶次，训练语音长度及辨识语音长度对系统辨识性能的影响。
|
| 关 键 词： | 说话人辨识 平均频谱 高斯混合模型 |
Text independent speaker identification system based on multiple strategies |
| |
| Mou Xiaolong,Hu Qixiu,Wu Wenhu.Text independent speaker identification system based on multiple strategies[J].Journal of Tsinghua University(Science and Technology),1997(3). |
| |
| Authors: | Mou Xiaolong Hu Qixiu Wu Wenhu |
| |
| Institution: | Mou Xiaolong,Hu Qixiu,Wu Wenhu Department of Computer Science and Technology,Tsinghua University, State Key Laboratory of Intelligent Technology and Systems,Beijing 100084 |
| |
| Abstract: | A speaker identification system, which can not only achieve high identification accuracy but also reduce the cost of calculation time and space, has been developed.The average spectrum feature is used in the first approach, followed by the corresponding decision rule.The Mel ceptral feature which is represented by Gaussian Mixture Model(GMM)is used in the second approach.A proof of the solving algorithm of GMM is given,and how the decision threshold,preemphasis coefficient,GMM order, utterance length for test and training affect the system performance has been studied. |
| |
| Keywords: | speaker identification average spectrum gaussian mixture model(GMM) |
| 本文献已被 CNKI 等数据库收录！ |
