Beyond Noise Suppression : Dynamics-driven Distortion Control Loss for Speech Enhancement and Robust Automatic Speech Recognition
- 주제(키워드) Speech enhancement , robust speech recognition , distortion control , noise suppression , loss function , Librispeech
- 발행기관 서강대학교 일반대학원
- 지도교수 Hyung-Min Park
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000082303
- UCI I804:11029-000000082303
- 본문언어 영어
- 저작권 논문은 저작권에 의해 보호받습니다.
목차
1 Introduction 1
2 Proposed Method 4
2.1 Speech Enhancement Model 4
2.2 Proposed Loss Function 5
2.2.1 Magnitude Distortion Control (Lmag) 6
2.2.2 Temporal Dynamics Preservation (Ldyn) 7
2.2.3 Residual Suppression (Lrs) 8
2.2.4 Final Loss Formulation 8
3 Experiments 10
3.1 Experimental Setup 10
3.1.1 Dataset 10
3.1.2 Training and Model Configuration 11
3.1.3 Baseline Loss Functions 12
3.2 Experimental Results 14
3.2.1 Signal-Level Performance Evaluation 15
3.2.2 Spectrogram Analysis 21
3.2.3 Impact of Various Loss Functions on ASR Performance 23
3.2.4 Generalization to Unseen SNR Conditions 24
3.2.5 Ablation Study on Loss Components 26
4 Conclusion 29
4.1 Summary 29
4.2 Limitations and Future Works 30
Bibliography 32
