Initial/final segmentation using loss function and acoustic features

To realize robust initial/final segmentation and satisfy the requirements of large vocabulary continuous speech recognition,an initial/final segmentation method that establishes the loss function and employs the quasiperiodicity of voiced sounds and the duration of initials and finals is proposed.The method firstly calculates the autocorrelation function of the speech,secondly establishes the loss function and performs voiced detection upon the results using the dynamic programming method,thirdly determines the duration of the initials according to their distributing rules,and finally segments the initials and finals from the two parts abutting on the beginning edge of the voiced region within the detection scope using the auditory event detection method.Experimental results show that the dynamic programming method is more efficient than the threshold method in detecting the voiced regions,the segmentation accuracy is improved as the segmentation is performed upon the voiced regions,the impacts of noises in speech and sound changes of Chinese are reduced,and the segmentation efficiency is seldom affected by the pronunciation styles of the initials.