北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2018, Vol. 41 ›› Issue (2): 56-61.doi: 10.13190/j.jbupt.2017-157

• 论文 • 上一篇    下一篇

基于谱信息熵和互补模型的声效模式检测

晁浩, 鲁保云, 刘永利, 刘志中, 宋成   

  1. 河南理工大学 计算机科学与技术学院, 河南 焦作 454000
  • 收稿日期:2017-08-10 出版日期:2018-04-28 发布日期:2018-03-17
  • 作者简介:晁浩(1981-),男,讲师,E-mail:chaohao1981@163.com.
  • 基金资助:
    国家自然科学基金项目(61502150,61403128);河南省高等学校青年骨干教师资助项目(2015GGJS-068);河南省科技攻关项目(172102210279)

Vocal Effort Detection Based on Spectral Information Entropy and Complementary Models

CHAO Hao, LU Bao-yun, LIU Yong-li, LIU Zhi-zhong, SONG Cheng   

  1. School of Computer Science and Technology, Henan Polytechnic University, Henan Jiaozuo 454000, China
  • Received:2017-08-10 Online:2018-04-28 Published:2018-03-17

摘要: 提出了一种基于模型融合的声效检测方法.首先提取对所有声效模式都具有良好辨识能力的谱信息熵特征,并进行声效辨识度分析;然后引入互补模型进行声效识别,从模型层面实现了整体谱特征、梅尔频率倒谱系数和谱信息熵的融合.对孤立词测试集进行了声效检测实验,识别精度为81.6%,实验结果表明,谱信息熵在3类特征中具有最好的分类能力,而互补模型能够有效集成3种特征蕴含的显著性信息.

关键词: 声效, 谱信息熵, 支持向量机, 高斯混合模型, 多层感知器

Abstract: A new vocal effort detection method based on model fusion was presented. By analyzing the ability to discriminate the vocal effort modes, the spectral information entropy feature which contains more salient information regarding the vocal effort level was proposed. Then, the complementary models were presented to achieve the fusion of the spectrum features, Mel-frequency cepstral coefficients and spectral information entropy feature. Experiments are conducted on isolated words test set, and the proposed method achieves 81.6% average recognition accuracy. The results show the spectral information entropy has the best performance among the three kinds of features and the complementary models can integrate the three kinds of features effectively.

Key words: vocal effort, spectral information entropy, support vector machine, Gaussian mixture model, multilayer perceptron

中图分类号: