基于HMM和ANN混合模型的语音情感识别研究  

Research on Speech Emotion Recognition Based on HMM and ANN Mixed Model

作　　者：林巧民[1,2] 齐柱柱 LIN Qiao-min;QI Zhu-zhu(School of Computer Science,Nanjing University of Posts ＆Telecommunications,Nanjing 210023,China;School of Educational Science and Technology,Nanjing University of Posts ＆Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学计算机学院,江苏南京210023 [2]南京邮电大学教育科学与技术学院,江苏南京210003

出　　处：《计算机技术与发展》2018年第10期74-78,共5页Computer Technology and Development

基　　金：国家自然科学基金(61572260);江苏省重点研发计划(BE2015702)

摘　　要：随着情感计算成为人工智能的一个重要方向,语音情感识别作为情感计算的一个重要部分,已经逐渐成为模式识别领域研究的热点之一。随着研究的不断深入,单独使用某一种模式识别时效果并不理想。为了提高识别率,提出了一种将隐马尔可夫模型(HMM)和径向基函数神经网络(RBF)相结合的方法。这种方法对不同情感状态分别设计HMM模型,经过维特比(Viterbi)算法得到最优状态序列,然后对得到的状态序列进行时间规整,以便生成等维的特征矢量,将其作为RBF模型的输入进行语音情感识别,最后的识别结果由RBF模型给出。实验结果表明,与孤立HMM相比,该方法在识别率上有较大的提高。As emotion calculation becomes an important direction of artificial intelligence,speech emotion recognition,as an important partof emotional computing,has gradually become one of the hot spots in the field of pattern recognition. With the development of the research,the recognition effect is not very ideal when just used a single model to classify speech emotional status. In order to improve recognition rate,we propose a method in combination of Hidden Markov Model （HMM） and radial basis function neural network （RBF）.This method designs HMM models for different emotional states,then gets the best sequence of emotional speech signal by Viterbi algorithm. Then,the feature parameters of the same state are structured as uniform dimension by the method of spatial orthogonal basis function expansion,which is used as the input of RBF for recognition of speech emotional states. Finally,the final results are given by RBF.The experiment shows that the proposed method has better recognition rate than isolated HMM.

关 键 词：情感计算 人工智能 隐马尔可夫模型 神经网络 语音情感识别 

分 类 号：TN912.34[电子电信—通信与信息系统]