Baum-Welch algorithm most likely results in underflow in practice. In some literatures, such as 'Scaling' algorithm was introduced to solve the problem. In applications, however, some mistakes were found in th...Baum-Welch algorithm most likely results in underflow in practice. In some literatures, such as 'Scaling' algorithm was introduced to solve the problem. In applications, however, some mistakes were found in the equations presented in these literatures. The practical calculations show that the original algorithm often results in poor or even none convergence and rather higher error rate in speech recognition. The mistakes in these literatures and brings forward the correct equations are analysed. The speech recognition system using the revised algorithm can converge well and has lower error rate.展开更多
Emotion recognition has become an important task of modern human-computer interac- tion. A multilayer boosted HMM （ MBHMM ） classifier for automatic audio-visual emotion recognition is presented in this paper. A mod...Emotion recognition has become an important task of modern human-computer interac- tion. A multilayer boosted HMM （ MBHMM ） classifier for automatic audio-visual emotion recognition is presented in this paper. A modified Baum-Welch algorithm is proposed for component HMM learn- ing and adaptive boosting （AdaBoost） is used to train ensemble classifiers for different layers （cues）. Except for the first layer, the initial weights of training samples in current layer are decided by recognition results of the ensemble classifier in the upper layer. Thus the training procedure using current cue can focus more on the difficult samples according to the previous cue. Our MBHMM clas- sifier is combined by these ensemble classifiers and takes advantage of the complementary informa- tion from multiple cues and modalities. Experimental results on audio-visual emotion data collected in Wizard of Oz scenarios and labeled under two types of emotion category sets demonstrate that our approach is effective and promising.展开更多
文摘Baum-Welch algorithm most likely results in underflow in practice. In some literatures, such as 'Scaling' algorithm was introduced to solve the problem. In applications, however, some mistakes were found in the equations presented in these literatures. The practical calculations show that the original algorithm often results in poor or even none convergence and rather higher error rate in speech recognition. The mistakes in these literatures and brings forward the correct equations are analysed. The speech recognition system using the revised algorithm can converge well and has lower error rate.
基金Supported by the National Natural Science Foundation of China(60905006)the NSFC-Guangdong Joint Fund(U1035004)
文摘Emotion recognition has become an important task of modern human-computer interac- tion. A multilayer boosted HMM （ MBHMM ） classifier for automatic audio-visual emotion recognition is presented in this paper. A modified Baum-Welch algorithm is proposed for component HMM learn- ing and adaptive boosting （AdaBoost） is used to train ensemble classifiers for different layers （cues）. Except for the first layer, the initial weights of training samples in current layer are decided by recognition results of the ensemble classifier in the upper layer. Thus the training procedure using current cue can focus more on the difficult samples according to the previous cue. Our MBHMM clas- sifier is combined by these ensemble classifiers and takes advantage of the complementary informa- tion from multiple cues and modalities. Experimental results on audio-visual emotion data collected in Wizard of Oz scenarios and labeled under two types of emotion category sets demonstrate that our approach is effective and promising.
