Target speech extraction based on auxiliary ICA with distortionless constraint for robust speech recognition
- 주제(키워드) 음성인식 , 음성추출
- 발행기관 서강대학교 일반대학원
- 지도교수 박형민
- 발행년도 2019
- 학위수여년월 2019. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000063903
- UCI I804:11029-000000063903
- 본문언어 영어
- 저작권 서강대학교 논문은 저작권보호를 받습니다.
초록/요약
This paper describes target speech extraction method for robust speech recognition. Signal obtained from microphone is often distorted with added noise comes from surrounding environment, which results in degradation of recognition performance. Therefore, target speech extraction method to extract target speech from noisy input is necessary for robust speech recognition. Conventional source separation methods are applicable to target extraction. Drawbacks of conventional methods, however, degrade recognition performance when it comes to target extraction problem. Frequency-domain Independent Component Analysis (FDICA), which is typical source separation method, suffers from problems, e.g. permutation, trade-off between learning rate and convergence speed, and scale indeterminacy of demixing matrix. This paper proposes target speech extraction method that deals with these drawbacks for robust speech recognition, using constraints on prior knowledge of target direction.
more