论文部分内容阅读
将基于多个嵌入图组合形式的半监督判别分析(SDA)以及核SDA(KSDA)应用于全监督的语音情感识别.在语音信号样本情感成分的预处理阶段,从样本语段中提取出多种特征及其统计参数,包括基音、过零率、能量、持续长度、共振峰和MFCC(Mel频率倒谱系数).在将样本特征送入分类器之前的维数约简阶段,使用经过参数优化的SDA或KSDA进行降维.Berlin语音情感数据库上的实验表明,在使用多类SVM分类器时的全监督语音情感识别中,SDA优于其他一些先进的基于谱图学习的维数约简算法,如LDA,LPP,MFA等,而KSDA通过核化的数据映射,能够取得比上述所有算法更好的识别效果.
The semi-supervised discriminant analysis (SDA) and the nuclear SDA (KSDA) based on the combination of multiple embedding images are applied to the full supervision of the voice emotion recognition. In the preprocessing stage of the emotional components of the voice signal samples, Including the pitch, zero-crossing rate, energy, duration, formant, and MFCC (Mel Frequency Cepstral Coefficient). In the dimension reduction stage before sample features are fed into the classifier, Optimized SDA or KSDA.Experiments on Berlin voice emotion database show that SDA is superior to other advanced spectral reduction based on spectral learning in full-supervised speech emotion recognition using multiple SVM classifiers Algorithms such as LDA, LPP, MFA and so on, and KSDA through nuclear data mapping, to achieve better recognition than all the above algorithms.