VMD改进GFCC的情感语音特征提取

Feature extraction of emotional speech based on improved GFCC with VMD

传统特征提取忽略了语音信号的非稳态特性,变分模态分解技术可以精细刻画语音的非平稳性,因此利用该技术将情感语音信号分解为K个固有模态函数,对每个分量做快速傅里叶变换后进行频率合成,通过Gammatone滤波器取能量对数,经离散余弦变换得到新特征变,即分模态分解改进Gammatone频率倒谱系数。通过支持向量机进行语音情感识别,实验结果表明,TYUT2.0中的识别率为72.15%,柏林情感语音库中的识别率为91.10%,识别效果优于传统情感语音特征,验证了该特征的有效性。

The traditional feature extraction ignores the non-stationarity of the speech signal,and the variational mode decomposition(VMD)technique can finely describe the non-steady state of speech.Therefore,a feature of improved Gammatone frequency cepstral coefficients based on VMD(VGFCC)was proposed.The steps of extracting VGFCC included decomposing signal into K intrinsic mode function(IMF)components using VMD,implementing fast fourier transform(FFT)for each IMF,frequency synthesis,Gammatone filter,logarithm,discrete cosine transform(DCT)and calculating statistical parameters.The extracted VGFCC was used to identify different emotions through the support vector machine(SVM).Experimental results show that VGFCC is superior to traditional emotional features,the recognition rate in TYUT 2.0 is 72.15%,and the recognition rate in EMO-DB is 91.10%.So the validity of the presented VGFCC is demonstrated.