摘要
为提高情感语音识别的正确率,研究了声学参数的统计特征和时序特征在区分情感中的作用,并提出了一种将两者相融合的情感识别方法。在提取出基本的韵律参数和频谱参数后,首先利用PNN(probab ilistic neura l netw ork)和HMM(h idden m arkov m ode l)分别对声学参数的统计特征和时序特征进行处理。计算它们各自属于每类情感的概率,获得采用加法规则和乘法规则融合统计特征和时序特征的识别结果。实验结果表明:各组特征在区分情感方面的侧重不尽相同,通过特征融合,平均识别正确率相较单独采用统计特征或时序特征均有提高,在最好情况下达到了92.9%。这说明了该方法的有效性。
A speech emotion recognition algorithm was developed based on the statistical and temporal fealures of the acoustic parameters for discriminating between emotions. The system first extracted the basic prosody parameters and spectral parameters, then. used a PNN (probabilistic neural network) to model the statistic features and a HMM (hidden Markov model) to model the temporal features. The sum and product rules were used to combine the probabilities from each group of features for the final decision. Experiments on the Cbinese speech corpus showed how the statistical and temporal features tend to reflect different aspects of emotions. The accuracy rate obtained by feature combination is higher than that by each group alone, reaching a maximum of 92.9%.
出处
《清华大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2006年第1期86-89,共4页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金资助项目(60433030
60418012)
关键词
语言识别
模式识别
情感信息处理
声学特征
speech recognition
pattern recognition
emotion information processing
acoustic features