Speaker recognition in an emotionalized spontaneous speech using empirical mode decomposition.

Speaker recognition in emotionalized spontaneous speech is the key technique of human-centered computing. Existed speaker recognition systems are hard to work during real-time application, because the extracted features of voiceprint aren't consistent in each time. This is caused by the variety of emotionalized intonation and contaminated acoustic signals of speaker. This is the problem we encountered when we tried to identify the speaker with the emotionalized spontaneous speech in real-time. To solve this problem, some inherent features from varying emotionalized spontaneous speech through different experiments are presented in the first part of this paper. Then the original speech signal is decomposed into several intrinsic mode functions based on the empirical mode decomposition. According to those intrinsic mode functions, the inherent features of the speaker's voiceprint are identified. Thus, a Gaussian mixture model can be constructed according to those features. However, the experiment based on our corpus of mandarin spontaneous speech showed one conclusion that the identifying rate is improved for fifteen percentage in average under our endeavors. [ABSTRACT FROM AUTHOR]