199 0

Speech Prediction using a Time-Varying Volterra Filter with Adaptive Wavelet-Basis FWLS Scheme

Title
Speech Prediction using a Time-Varying Volterra Filter with Adaptive Wavelet-Basis FWLS Scheme
Other Titles
Wavelet-basis FWLS 알고리즘과 시변 Volterra filter를 이용한 적응 음성 예측 기법
Author
이영일
Alternative Author(s)
Lee, Young-Il
Advisor(s)
남상원
Issue Date
2007-02
Publisher
한양대학교
Degree
Master
Abstract
음성 예측(speech prediction) 기법은 과거의 음성 신호 값들을 이용하여 미래의 신호(sample)를 예측하는 기법으로, 인체의 조음 기관을 효과적으로 모델링하여야 그 구현이 가능하며, 신호처리의 다양한 분야에서 응용되고 있다. 음성 예측기(speech predictor)는 조음 기관이 갖는 비선형(nonlinear) 성질과 시변(time-varying) 성질로 인하여 발생하는 예측 오차(prediction error)를 줄이는 것이 핵심적인 문제이며, 이를 해결하기 위하여 많은 알고리즘들이 개발되어 왔다. 본 논문에서는 wavelet basis FWLS 기법을 적응 알고리즘으로 하는 Volterra 필터를 이용하여 음성 예측을 함으로써 음성 신호가 갖는 시변 비선형 성질을 모델링하여 예측 오차를 효과적으로 줄일 수 있음을 실험 데이터를 통하여 검증하였다.; The linear prediction (LP) model technique is a very important method in digital signal processing. Especially, speech prediction has been used in diverse applications, such as speech coding, restoration, noise reduction, etc. However, even though methods have been developed for speech prediction, it is difficult to reduce do not reduce the prediction error more efficiently due to the time-varying nature of the voice track and non-linearities in the speech signals. In this thesis, a new adaptive nonlinear predictor is proposed, whereby time-varying Volterra filtering with a wavelet-basis FWLS (filtered weighted least squares) algorithm is employed to model properly the time-varying characteristics and non-linearities of speech signals. Simulation results demonstrate that the proposed approach yields better convergence and much reduced error performance for speech prediction, compared with conventional methods.
URI
https://repository.hanyang.ac.kr/handle/20.500.11754/150010http://hanyang.dcollection.net/common/orgView/200000405546
Appears in Collections:
GRADUATE SCHOOL[S](대학원) > DEPARTMENT OF ELECTRONICS & COMPUTER ENGINEERING(전자통신컴퓨터공학과) > Theses (Master)
Files in This Item:
There are no files associated with this item.
Export
RIS (EndNote)
XLS (Excel)
XML


qrcode

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE