Entropy-based initial/final segmentation for Chinese whiskered speech

The Initial/Final(IF) segmentation of whispered speech is the pre-processing in the whispered speech recognition and the reconstruction of normal speech from whisper. However, because the whispered initials and finals are all unvoiced, it is difficult to segment them by the methods used in the normal speech. With tile characteristics analysis of Chinese whispered speech, a new segmentation method is proposed. The speech endpoint is detected by the entropy function, and the initial/final boundary is obtained by the decision of the initial duration, the symmetric relative entropy and the normalized spectral center of gravity. The correct segmentation rates are 87.9% for the female data and 90.3% for the male data in the test with 380 Chinese whispered syllables at 2-10 dB SNR. It is more accuracy than the frequency domain method, the clustering method and the spectral flatness method. As shown in the experiments, this algorithm can be used as pre-processing in the whispered speech recognition and the conversion. It gives the reconstructed speech a more natural quality.