KR930008721A

KR930008721A - Feature Extraction Method of Speech Recognition System

Info

Publication number: KR930008721A
Application number: KR1019910018048A
Authority: KR
Inventors: 김탁용
Original assignee: 이헌조; 주식회사 금성사
Priority date: 1991-10-14
Filing date: 1991-10-14
Publication date: 1993-05-21
Also published as: KR0176751B1

Abstract

본 발명은 음성인식 시스템의 특징 추출방법에 관한 것으로, 종래에는 배경소음이 음성신호와 같이 입력되는 경우 기준패턴과의 차이에 의해서 음성인식 시스템의 성능이 떨어지므로, 본 발명은 소음 존재시 소음이 없는 조용한 환경에서 만들어진 표준패턴의 차이를 선행정규화 방법을 이용해서 음원신호의 크기변화에 따른 영향은 물론 소음에 대한 영향을 거의 받지 않게 함으로써 인식시스템의 성능을 개선시키도록 한 것이다.The present invention relates to a feature extraction method of a speech recognition system, and in the related art, when the background noise is input as a speech signal, the performance of the speech recognition system is reduced by a difference from a reference pattern. The difference of the standard pattern made in the quiet environment is improved by using the prior normalization method to improve the performance of the recognition system by minimizing the effects of the change of the size of the sound source signal and the noise.

즉, 대역필터와 저역필터 및 정규화 과정으로 구성되는 특징인자 추출방법에 관한 것으로, 정규화 과정으로 구성되는 특징인자 추출방법에 관한 것으로, 정규화과정에서 선형에너지를 사용하는 점과 그 선형에너지를 평균화 및 최대값으로 나누는 방법으로 첨가소음에 대해 그 영향을 제거함으로써, 준 백색성의 잡음(소음)이 존재하는 실제환경에서 사용되는 음성인식 시스템의 특징추출과정에 중요한 용도로 쓰이게 된다.That is, the present invention relates to a feature factor extraction method comprising a band pass filter, a low pass filter, and a normalization process, and a feature factor extraction method comprising a normalization process. By dividing the maximum value by removing the effect on the added noise, it is used for the feature extraction process of the speech recognition system used in the real environment where quasi-white noise (noise) exists.

Description

Feature Extraction Method of Speech Recognition System

본 내용은 요부공개 건이므로 전문내용을 수록하지 않았음Since this is an open matter, no full text was included.

제 4 도는 본 발명에 의한 선형에너지를 이용한 정규화 방법을 보인 신호흐름도,4 is a signal flow diagram showing a normalization method using linear energy according to the present invention,

제 5 도의 ㈎ 내지 ㈑는 본 발명에 의한 클린 및 노이즈 스피치인 각 경우의 선형과 대수 스펙트럼을 보인 파형도,Fig. 5 is a waveform diagram showing linear and logarithmic spectra in each case of clean and noise speech according to the present invention;

제 7 도 ㈎ 내지 ㈓는 본 발명에 의한 선형정규화 과정의 각 단계에서의 기준패턴과 테스트패턴의 스펙트럼 비교를 보인 파형도.7 is a waveform diagram showing a spectral comparison between a reference pattern and a test pattern at each step of the linear normalization process according to the present invention.

Claims

The digitally converted voice signal S (n) is band-filtered and nonlinearly calculated by the band filter units 9A-9N and the nonlinear operation units 10A-10N, and then the low-pass filter units 11A-11N and the switches ( In the speech recognition system which calculates the logarithm and linear energy of each channel by SW1-SW3) and performs n-dimensional feature vectors in the normalization process, the initial state of the current channel number (i) is 1 and sum (SUM) is 0. If the number of channels (i) is equal to or greater than the set number of channels (n) by adding the linear energy {xi (m)} of the reference pattern to the sum, the channel average value due to the linear energy And the mean value {x (m)} If the linear energy {xi (m)} subtracted from is the maximum value and the current channel (j) is less than or equal to the set number of channels (n), the linear energy {xi (m)} from which the average value is removed is taken as the maximum value (Max). A method for extracting features of a speech recognition system, characterized in that it is represented by the calculated and normalized linear energy {xi (m)}.

※ Note: The disclosure is based on the initial application.