关于人的语音声调准确识别仿真
发布时间：2018-12-17 17:55
【摘要】：人的语音声调的准确识别,可以提高语音信号处理效果,保证人机通信的顺利进行。声调识别时,需要获取不同的声调模式,将待识别的声道进行比对,而传统的基于RNN-RBM语言模型的识别方法只能获取语音音素、单词以及语句,不能获取其对应的标准声调模式,无法完成比对,降低了识别的精度。提出基于K-means初始化EM算法的语音声调识别方法。通过建立声调信息高斯混合模型,准确的对基频信息概率密度函数进行拟合,采用最大化(EM)算法提取基频特征参数,并以此为基础获取更多的声调模式,利用K-means初始化EM算法,消除EM算法对初始值选取较为敏感的问题,再对高斯混合模型阶数进行预测,提高EM算法执行声调识别精度。仿真结果表明,采用改进的声调识别方法进行声调识别,识别准确率较高,具有一定的实用性。
[Abstract]:The accurate recognition of human voice tone can improve the effect of speech signal processing and guarantee the smooth progress of communication. In tone recognition, different tone patterns need to be obtained and the tracks to be recognized are compared. However, the traditional recognition method based on RNN-RBM language model can only obtain phoneme, word and sentence. The accuracy of recognition can not be reduced because the corresponding standard tone mode can not be obtained and the comparison can not be completed. A speech tone recognition method based on K-means initialized EM algorithm is proposed. By establishing the mixed model of tone information Gao Si, the probability density function of fundamental frequency information is fitted accurately, and the feature parameters of fundamental frequency are extracted by maximization (EM) algorithm, and more tone modes are obtained based on this model. Using K-means to initialize EM algorithm, the problem that EM algorithm is sensitive to initial value selection is eliminated, and then the order of Gao Si mixed model is predicted to improve the accuracy of tone recognition of EM algorithm. The simulation results show that the improved tone recognition method has high accuracy and practicability.
【作者单位】： 甘肃农业大学信息科学与技术学院;
【分类号】：TN912.34
[Abstract]:The accurate recognition of human voice tone can improve the effect of speech signal processing and guarantee the smooth progress of communication. In tone recognition, different tone patterns need to be obtained and the tracks to be recognized are compared. However, the traditional recognition method based on RNN-RBM language model can only obtain phoneme, word and sentence. The accuracy of recognition can not be reduced because the corresponding standard tone mode can not be obtained and the comparison can not be completed. A speech tone recognition method based on K-means initialized EM algorithm is proposed. By establishing the mixed model of tone information Gao Si, the probability density function of fundamental frequency information is fitted accurately, and the feature parameters of fundamental frequency are extracted by maximization (EM) algorithm, and more tone modes are obtained based on this model. Using K-means to initialize EM algorithm, the problem that EM algorithm is sensitive to initial value selection is eliminated, and then the order of Gao Si mixed model is predicted to improve the accuracy of tone recognition of EM algorithm. The simulation results show that the improved tone recognition method has high accuracy and practicability.
【作者单位】： 甘肃农业大学信息科学与技术学院;
【分类号】：TN912.34
【参考文献】
相关期刊论文 前7条
1 谭萍;邢玉娟;;噪声环境下文本相关说话人识别方法改进[J];西安工程大学学报;2016年05期
2 张毅;黎小松;罗元;吴承军;;基于人耳听觉特性的语音识别预处理研究[J];计算机仿真;2015年12期
3 高霞;李瑞俊;;EM算法在不完全数据参数估计中的应用[J];集宁师范学院学报;2015年03期
4 张玉梅;胡小俊;吴晓军;白树林;路纲;;语音信号序列的Volterra预测模型[J];物理学报;2015年20期
5 吴蔚澜;蔡猛;田W，
本文编号：2384592
本文链接：https://www.wllwen.com/kejilunwen/xinxigongchenglunwen/2384592.html
