Multi-Channel Speech Enhancement using Beamforming and Nullforming for Severely Adverse Drone Environment
- 발행기관 서강대학교 일반대학원
- 지도교수 박형민
- 발행년도 2024
- 학위수여년월 2024. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000076932
- UCI I804:11029-000000076932
- 본문언어 영어
- 저작권 서강대학교 논문은 저작권 보호를 받습니다.
초록
Recent end-to-end neural beamforming have shown impressive results in multi-channel speech enhancement tasks. Particularly, EaBNet, integrating embedding and beamforming techniques, addresses the lack of interpretability inherent in previous end-to-end beamforming models. This model utilizes spatial information for speech enhancement and seeks further improvement through its PostNet. However, in environments like flying drones, the noise from propellers and motors is closer and louder compared to the target speech, creating low SNR conditions. In such severely adverse drone environment, residual noise remains even after beamforming, highlighting the importance of PostNet. Drawing inspiration from the structure of the Generalized Sidelobe Canceller (GSC) algorithm, this paper proposes enhancing the performance of the PostNet by estimating the spatial information of ego-noise through nullforming. Additionally, it aims to facilitate the training of the PostNet by estimating the spectral features of the speech and noise estimation outputs.
more목차
1 Introduction
1.1 Background
1.2 Overview of the proposed method
1.3 Outline
2 Signal Model and Traditional Beamforming
2.1 Signal Model
2.2 Beamforming
2.2.1 MVDR Beamforming
2.2.2 MaxSNR Beamforming
2.2.3 MaxNSR Beamforming
3 EaBNet
3.1 Network structure
3.2 Embedding and Beamforming Module
3.3 PostNet
4 Proposed Method
4.1 Another Algorithm for Beamforming Module’s Target
4.2 Double BM
4.3 Additional Spectral Feature Extraction
5 Experimental Setup
5.1 Dataset Preparation
5.2 Model Details
5.3 Training Details
6 Results and Discussions
6.1 Result of the Proposed Method
6.2 Discussions
7 Conclusion
Bibliography
