基于高斯混合模型的情感LPC系数的研究

Research on emotional LPC parameter based on GMM

针对当前缺乏语音情感特征的发声模型研究的具体现状,通过探索情感特征与线性预测编码(LPC)系数之间的映射关系,提出建立LPC系数情感模型的新方法。该方法在中科院情感语音库的基础上分别建立高兴、愤怒、悲伤及中性4种情感数据库;通过采用不同情感语音的LPC特征矢量,结合动态时间规整技术(DTW)、期望最大化算法(EM)和最小均方误差准则(MMSE)对高斯混合模型进行训练和参数估计,最终获得高兴、愤怒、悲伤这3种情感对中性语音的LPC系数映射规则函数,完成对情感LPC系数的建模。研究结果表明:建立的情感模型有效表征了不同情感对LPC系数的影响;该方法为情感语音合成、识别等研究提供了新思路。

Considering the inadequacy of the study on emotional speech modeling,a new method of emotional speech modeling of linear predictive coding(LPC) parameter was presented.The method established four kinds of emotional speech database including happy,angry,sad and neutral emotion recorded by Chinese Academy of Sciences.Different emotional LPC feature vectors were used combined with the dynamic time warping technology,expectation maximization(EM) algorithm and minimum mean square error(MMSE) criterion.Finally LPC parameter mapping rule function of three kinds of emotional speech to neutral speech was gotten.The emotional LPC parameter modeling was completed.The results show that the emotional speech model can efficiently characterize the different emotional effects on LPC parameter.The new method provides a new idea to the research of the influence for emotional speech synthesis and recognition.