협대역 음성의 광대역 변환을 위한 신경망 기반의 잔차신호 복원 = Reconstruction of Residual Signal based on Neural Networks For Narrowband-to-Wideband Speech Conversion.

다국어 초록 (Multilingual Abstract)

This paper investigates the use of neural network for narrowband to wideband speech conversion that can be used for speech quality enhancement in the telephone communications. Neural network has a good ability to approximate a function. So it is suitable for continuous mapping from narrowband to wideband speech. this paper propose two single neural networks for speech enhandcement. one is spectral envelope expansion, another is wideband excitation generation. but this paper concentrate heavily on wideband excitation generation than spectral envelpe expansion. Experimental results show the effectiveness of the proposed methods.

목차 (Table of Contents)

• 목차
• 1. 서론 = 1
• 2. 기존의 음성 대역폭 확장방법 = 4
• 2.1 Codebook mappping 방법을 이용한 spectral envelope 확장 = 6
• 2.2 Codebook mapping 문제점 = 7
• 3. Multilayer perceptron (MLP) = 8
• 3.1 MLP 네트워크 구조 = 8
• 3.2 Back-Propagation Algorithm = 10
• 4. 제안 방법 = 14
• 4.1 신경망으로 복원된 여기신호를 이용한 Narrowband 음성에서의 Wideband 음성의 변환 = 14
• 4.2 Wideband Excitation Generation using SNN = 16
• 5. 실험 방법 = 18
• 5.1 Narrowband / wideband residual cepstrum 추출 = 18
• 5.2. BP training & test using extracted residual cepstrum = 21
• 5.3 Residual generation = 24
• 5.4 Wideband speech generation using resconstructed wideband residual = 26
• 6. 모의 실험 결과 = 28
• 7. 결론 = 32
• 참고문헌 = 34
• Abstract = 36