关键词: 声韵分割, 经验模态分解（EMD）, 语谱图

Abstract: Accurate initial-final segmentation is considered as a basis and key work for Chinese speech recognition and synthesis etc. Since short-time energy and zero cross rate which are used in traditional segmentation are easily disturbed by noises and cause further segmentation error, spectrogram is introduced here to Chinese monosyllable initial-final segmentation due to its time-frequency information. It is proved to be effective by experiment, it is then combined with empirical mode decomposition and keeping-sign-ratio computing in Chinese dual-syllable initial-
final segmentation. As is shown in experiment, the proposed time-frequency method has achieved an accuracy of 86.92% and 77.47% in the initial-final segmentation for Chinese monosyllable and dual-syllable respectively.

Key words: initial-final segmentation, Empirical Mode Decomposition（EMD）, spectrogram