郑州大学学报（理学版）

针对隐马尔科夫模型和径向基神经网络识别语音情感的缺陷,提出了一种新的基于两者的混合模型识别方法.将神经预测器引入隐马尔科夫模型计算状态观察概率,使得隐马尔科夫模型能有效利用语音帧间信息,同时又利用状态累计概率输入径向基神经网络分类,避免特征向量时间规整的麻烦.选取高兴、愤怒、悲伤、惊奇、平静五类情感,以平静情感为参照求取特征向量进行实验,结果表明,相对于单一隐马尔科夫模型和常用混合模型方法,该混合模型识别性能有明显改善,并且加入噪声后的识别效果仍然较好.

According to the analysis of the emotional speech signal,a hybrid model for speech emotion recognition,consisting of a neural network with radial basis functions and hidden Markov models,was described.Because radial basis functions could describe correlation between the frames,this hybrid model utilized a neural net with radial basis functions to approximate posterior probabilities of hidden Markov models states.In addition,the discriminating recognition results were recognized by radial basis functions.In the experiments,five kinds of emotional speech such as happy,anger,sadness,surprise,and calm were recorded.The relative values of the first four kinds of emotional characteristic parameters and calm emotional characteristic parameters were regarded as the input characteristic vector.The experiments showed that the improved hybrid model with additional noise to the signal was better than hidden Markov models and the pre-existing hybrid model.