Speech endpoint identification based on empirical mode decomposition

A new method based on the empirical mode decomposition (EMD) was proposed to identify speech-segment endpoints in noise-contaminated speech signals. Noisy speech signals were decomposed into a set of intrinsic mode functions (IMFs) using EMD. The average instantaneous frequencies of IMFs were estimated by their short time zero cross rate. The frames with low and slowly changing average instantaneous frequencies were identified to be the periodic sonant segments and the frames with high average instantaneous frequencies were identified to be the surd segments based on the characteristics of the average instantaneous frequencies of IMFs derived from speech signals. The final speech signals were obtained by processing and combining these segments. The numerical and experimental results show that the method can effectively identify the endpoints for the speeches contaminated by noises seriously.