基于语音特征的情感分类  

Sentiment Classification Based on Voice Features

作　　者：李强[1,2] 刘晓峰[3] 贺静[4] 

机构地区：[1]太原理工大学财经学院信息系,太原030024 [2]山西省财政税务专科学校经济信息系,太原030024 [3]太原理工大学数学学院,太原030024 [4]太原理工大学信息化管理与建设中心,太原030024

出　　处：《小型微型计算机系统》2016年第2期385-388,共4页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61072087)资助

摘　　要：情感识别是智能机器人技术研究中的一个重要课题,对不同的情感识别有助于提高机器人的智能水平和与人类沟通的效果.首先采用语速、瞬时能量、瞬时过零率、共振峰和基频五种语音特征,以及正常、喜悦、愤怒、悲伤、惊讶五种情绪状态,建立语音特征与情绪的相关性模型,然后设计PNN识别算法,再通过该算法对情绪状态训练分类,最后在语音识别过程中提取的低阶特征,识别高阶情感语义.通过实验效果对比分析,其平均识别率可以达到82.58%,优于HMM的82.2%,主成分析的66.16%和多元回归分析的62.68%,可以得出该模型对语音情感识别有较好的识别率.Emotion recognition is an important research topic in the intelligent robot technology, and help to improve the intelligent lev- el of the robot and the effect of communicating with human. In the first place, this paper analyzes the five kinds of voice features, in- cluding instantaneous energy,instantaneous speed,zero crossing rate, fundamental frequency resonance peak）, as well as five kinds of mood, （ including normal, joy, anger, sadness, surprise ）, to establish the correlation model of emotional voice features. In the next place, we design PNN recognition algorithm, and train it by the emotional state for the classification. Finally the low order feature is extracted and high order emotion semantic is identified in the process of voice recognition. Through the comparative analyses of experi- ments effects,the average recognition rate can reach 82. 58% ,that better than the 82. 2% of HMM,the 66. 16% of main analysis,and the 62. 68% of multiple regression analysis, so this model has good recognition rate of voice emotion recognition.

关 键 词：情感计算 语音识别 PNN 贝叶斯分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]