Emotion is communicated through a universal language that people instinctively know beyond their language, voice, face, race and so on. If people know emotion of others, they can look for more appropriate rational responses. The purpose of this resear...
Emotion is communicated through a universal language that people instinctively know beyond their language, voice, face, race and so on. If people know emotion of others, they can look for more appropriate rational responses. The purpose of this research is to help people communicate with machines naturally and emotionally.
We propose an improved speech emotion recognition algorithm which adopts temporal derivative features and support vector machines (SVMs). The proposed algorithm consists of feature extraction and SVM-based pattern classification. In the feature extraction module, features are extracted from speech signals such as pitch, jitter, mel-frequency cepstral coefficient (MFCC), energy, shimmer, and formant frequencies. In order to reflect long-term variation of speech signals, we append delta components of MFCC or shifted delta cepstrum (SDC) to the conventional features. Then, the functionals such as mean value, standard deviation, interquartile range are computed to obtain a fixed utterance-wise feature vector from a sequence of variable frame-wise feature vectors. In the pattern classification module, the model parameters of SVM are trained by using the training data set. In the test phase, the final emotion is decided by SVM.
In computer experiments, we use cross-validation where 9 speakers are used for training and the other 1 speaker for test. Emotion is classified into 4 classes: angry, neutral, sadness, and happiness. We evaluate and compare the performance of emotion recognition for two languages: German and Korean. We achieve 82.2% and 81.4% of emotion recognition accuracy for a public-domain German speech database and 55.4% and 54.0% for a Korean speech database using delta and SDC, respectively.
，免费韩语论文，韩语论文题目 |
