Precise detection of isolated words in noise by the linear prediction method

Abstract

This paper presents a new method of detection for speech in stationary noise. First, a fundamental principle of segmentation is derived. It is shown that the final prediction error (FPE) decreases monotonically as the length of the processed signal increases, if the signal is stationary or homogeneous, and the FPE tends to increase when the homogeneity of the signal is lost. This means that the instant when the FPE is at its minimum corresponds to the position of the initial word. Based on this principle, the final prediction error is proposed as a criterion to find the speech segments in noise precisely. Second, experiments to test the performance of the proposed method are presented. Experimental results show that the proposed method is sensitive enough to detect changes in signal characteristics. A detailed comparison of the accuracy of the proposed method and that of conventional methods has been made and the effectiveness of the proposed method is shown.