The Research of Endpoint Detection and Initial/Final Segmentation for Chinese Whispered Speech

This paper gives the suitable character parameters for endpoint detection and initial/final (I/F) segmentation of whispered speech based on waveform characteristics. At first, the whispered speech is analyzed to two parts approximation and detail - by discrete wavelet transform (DWT). Then fractal dimension of approximation part is calculated to be the character parameter of speech and non speech segmentation. And the energy ratio of the approximation and detail part (DAER) is calculated to be the parameter of initial and final segmentation. Finally, an efficient detection algorithm is given to achieve the endpoint detection and I/F segmentation. The experiments show that this is a simple and good efficiency algorithm, which is a suitable method for endpoint detection and UF segmentation of whispered speech