Emotion Classification via Modified Gaussian Mixture Model
Paper ID : 1744-IST
1Zeinab Hosseini *, 2Mohammad Ahadi, 3Neda Faraji
1Speech Processing Research Lab (SPRL)
Amirkabir University of Technology, Tehran, Iran
3Imam Khomeini International University
Abstract— Besides too much information that is concealed in speech signal, emotional state of the speaker is an important characteristic of the signal in speech research field. However, this problem is remained in negligible part in Automatic Speech Recognition (ASR). This study is focused on finding more effective methods to improve speaker emotional state classification .Two methods are proposed for training and test phases while the Gaussian Mixture Model (GMM) is selected as baseline system. In these methods the motivation is to reduce the confuse information region from emotion speech space and salience the discriminative region. In the training phase, symetric Kullback-Leibler Divergence (KLD) is used to detect the discriminative GMM mixtures while the confused region is ignored. This algorithm is known as KLD-GMM. In the test phase, the discriminative frames are recognized based on Frame Selection Decoding (FSD). This method is known as FSD-GMM algorithm. Two algorithms have led to an average improvement of about 7% in the emotion recognition system performance in comparison with generalized Gaussian Mixture Model which is utilized as baseline method.