Speech Dereverberation Using WPE Based on Complex Generalized Gaussian Distribution with Time-Invariant Full-Spatial Correlation Matrix
- 주제(키워드) Dereverberation , WPE
- 발행기관 서강대학교 일반대학원
- 지도교수 박형민
- 발행년도 2019
- 학위수여년월 2019. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000063862
- UCI I804:11029-000000063862
- 본문언어 영어
- 저작권 서강대학교 논문은 저작권보호를 받습니다.
초록/요약
In this paper, we propose a dereverberation algorithm for robust speech recognition in a reverberant environment. Recently, many products that implement auto speech recognition (ASR) system have been released. ASR performance degrades in reverberant environments. The importance of multi-channel preprocessing algorithm, including dereverberation algorithm, is growing. The conventional batch-processing weighted prediction error (WPE) perform well, but the size of the inverse matrix, which requires a lot of computation, is large. In this paper, we derive WPE based batch processing with the smaller size of the inverse matrix compare to conventional methods by rearrange prediction filter. Since clean speech has sparse distribution, we modeled dereverberated signals to the Generalized Gaussian distribution with a time-invariant full-spatial correlation matrix which showing performance improvement. In addition, we suggest time-varying Gaussian exponent, one of the Generalized Gaussian distribution component. We also propose a new full-spatial correlation matrix estimation equation that performs better than the originally derived equation.
more