Voice feature detection method using the convergence of the pulse voice source and pulse feature extraction by the entropy

Abstract

Background/Objectives: There are disadvantages that there should be a signal input separately for securing the noise signal reflecting the noise characteristics to remove the noise in the voice signal processing and performance degradation of low signal-to-noise ratio (SNR) occurs in a noisy environment. Methods/Statistical analysis: In this paper, the feature obtained from the original voice signal indicates the fundamental frequency of the voice signal and the representative section of the pulse voice source is restored for each feature cycle to generate a pulse voice source. We constructed the voice recognition model by extracting a pulse feature for each frame and proposed a voice feature detection method that combines energy spectrum entropy and pulse voice source. Findings: The proposed method extracts the features for classification of voiced and unvoiced at high signal-to-noise ratio (SNR) and constructs a model for recognition so that the characteristics of voice are less influenced by noise. Improvements/Applications:In the voice recognition, the excellent recognition rate was confirmed compared with the existing method and the recognition rate of approximately 2.45% P in the overall average for the voice dependent stage and the voice Independent stage was improved compared with the existing method. © 2018, Institute of Advanced Scientific Research, Inc.. All rights reserved.