Improved endpoint detection of multi-parameter fusion under noise reduction

Abstract

In order to overcome the low accuracy of traditional endpoint detection algorithms in low SNR speech endpoint detection. In this paper, EMD is used to decompose manic speech adaptively, and multi-window spectrum subtraction is used to reduce the noise of each decomposed signal to improve the SNR of speech signal. Based on the MFCC principle, the improved MFCC spectrum entropy was proposed, and then the product of logarithmic energy and MFCC spectrum entropy was taken as a new parameter, and median filtering was performed on it. Two dynamic thresholds were designed for this new feature parameter, which could better track and locate the start and end of the actual speech. Matlab simulation results show that compared with other common algorithms, the detection accuracy of the improved algorithm is improved by 2.2% under the -5dB Gaussian white noise, and by 2% under the -5dB Volvo noise. Not only from the perspective of innovation, but also from the perspective of performance, it has good accuracy and certain robustness under the low SNR environment.