摘要
同一组特征向量对不同的情感,其识别能力不同。以HMM作为语音情感分类器模型,对不同情感状态选择不同的特征向量进行识别。系统分两个阶段完成:首先基于漏识率和误识率最小的决策原则,采用优先选择(PFS)算法分别为每种情感状态选择最优的特征向量,然后用这些特征向量分别建立对应情感状态的HMM模型。利用北航情感语音库(BHUDES)对算法进行验证,将所有实验样本分为训练样本集、特征选择样本集和测试样本集3组,采用交叉实验的方法对本算法进行验证,结果表明,与单特征向量HMM相比,多特征向量HMM可达到更高的识别精度。
Same feature vector from speech may recognize different emotion state in different reliability. HMM was used as basic classifier and different feature vectors were chose as the input of HMM for different emotion. Firstly, based on the decision principle of the least miss-recognition rate and error-recognition rate, promising first selection (PFS) was adopted to choose the optimal feature vector for each emotion. Then, HMM for each emotion was set up using the selected feature vector. Cross experiments were implemented using Beihang University Database of emotion speech (BHUDES). All samples were divided into three groups: training sample set, feature selection sample set and test sample set. The experimental results show that HMM with multiple feature vectors can achieve better recognition precise than that with single feature vector.
出处
《计算机科学》
CSCD
北大核心
2009年第6期231-234,共4页
Computer Science
基金
国家863计划资助项目(2006AA01Z135)
教育部博士点基金资助项目(20070006057)资助
关键词
多特征向量
优先选择算法
决策
漏识率
误识率
Multiple feature vectors, Promising first selection (PFS), Decision, Miss-recognition rate, Error-recognition rate