지원사업
학술연구/단체지원/교육 등 연구자 활동을 지속하도록 DBpia가 지원하고 있어요.
커뮤니티
연구자들이 자신의 연구와 전문성을 널리 알리고, 새로운 협력의 기회를 만들 수 있는 네트워킹 공간이에요.
논문 기본 정보
- 자료유형
- 학위논문
- 저자정보
- 지도교수
- 정진우
- 발행연도
- 2016
- 저작권
- 상명대학교 논문은 저작권에 의해 보호받습니다.
이용수5
이 논문의 연구 히스토리 (2)
- 이 논문의 후속연구가 궁금하신가요?
- 연관 학술논문 또는 학술발표를 통해 보다 발전된 연구결과를 확인하실 수 있습니다.
- 이 논문의 연구 히스토리 확인하기
초록· 키워드
오류제보하기
전화 녹취 파일에서의 화자 구분
본 논문에서 제안하는 화자 구분(Speaker diarization) 방법은 다음과 같다. 우선 음성파일 (*.wav)에서 음성이 없는 구간(묵음구간)과 음성이 있는 구간(음성구간)을 분리하기 위해 음성파일의 앞 30초 구간의 Time Signal로 Envelop graph를 추출하여 묵음구간(silence)과 음성구간(non-silence)의 경계값(Threshold)을 추출한다. 두 번째로, 경계값(threshold)을 이용하여 전체 음성파일의 음성 구간을 분리하고, 2차에 걸친 세그멘테이션 작업을 한다. 세 번째로, 각각의 세그먼트를 FFT(Fast Fourier Transform)을 하여 Time signal을 Frequency signal로 변환한다. 네 번째로, FFT로 구한 Frequency signal에서 의미있는 주파수 특징 12개를 추출하기 위해 MFCC(Mel-Frequency Coefficient Cepstral)를 수행한다. 다섯 번째로, 12개의 MFCC 특징을 이용하여 각 Segment를 Gaussian Model로 만들고, 각 모델을 합하여 GMM(Gaussian Mixture Model)을 형성한다. 여섯 번째로, GMM을 기반으로 EM-Clustering을 수행하여 두 화자를 구분한다.
화자구분 결과의 오류율을 계산하기 위해 Ground truth로 음성 파일을 들으면서 실제 화자가 발화 하는 시간을 기록하여 알고리즘에 의해 나온 결과와 비교하였다. 그 결과 평균 5%이내의 DER(Diarization Error Rate)을 얻었다.
본 논문에서 제안하는 화자 구분(Speaker diarization) 방법은 다음과 같다. 우선 음성파일 (*.wav)에서 음성이 없는 구간(묵음구간)과 음성이 있는 구간(음성구간)을 분리하기 위해 음성파일의 앞 30초 구간의 Time Signal로 Envelop graph를 추출하여 묵음구간(silence)과 음성구간(non-silence)의 경계값(Threshold)을 추출한다. 두 번째로, 경계값(threshold)을 이용하여 전체 음성파일의 음성 구간을 분리하고, 2차에 걸친 세그멘테이션 작업을 한다. 세 번째로, 각각의 세그먼트를 FFT(Fast Fourier Transform)을 하여 Time signal을 Frequency signal로 변환한다. 네 번째로, FFT로 구한 Frequency signal에서 의미있는 주파수 특징 12개를 추출하기 위해 MFCC(Mel-Frequency Coefficient Cepstral)를 수행한다. 다섯 번째로, 12개의 MFCC 특징을 이용하여 각 Segment를 Gaussian Model로 만들고, 각 모델을 합하여 GMM(Gaussian Mixture Model)을 형성한다. 여섯 번째로, GMM을 기반으로 EM-Clustering을 수행하여 두 화자를 구분한다.
화자구분 결과의 오류율을 계산하기 위해 Ground truth로 음성 파일을 들으면서 실제 화자가 발화 하는 시간을 기록하여 알고리즘에 의해 나온 결과와 비교하였다. 그 결과 평균 5%이내의 DER(Diarization Error Rate)을 얻었다.
목차
차 례차례 ·················································································································· ⅰ표차례 ·············································································································· ⅱ그림차례 ·········································································································· ⅲ국문 요약 ········································································································ ⅴ1. 서론 ··············································································································· 12. 화자 구분에 대한 고찰 ············································································· 52.1. 화자 구분을 위한 단위 ········································································ 52.2. 음성의 특징 및 표현 ············································································ 62.3. DFT vs FFT ··························································································· 72.4. 클러스터링 기법 ·················································································· 153. 기존의 화자 구분 알고리즘 ··································································· 163.1. 화자 구분 알고리즘 ············································································ 163.2. 기존 화자 구분 알고리즘의 한계점 ················································ 184. 제안하는 화자 구분 알고리즘 ······························································· 194.1. 묵음 구간 검출 ···················································································· 214.2. 세그멘테이션 ························································································ 244.3. MFCC (Mel-Frequency Coeifficient Cepstral) ······························· 264.5. GMM (Gaussian Mixture Model) ······················································· 344.6. EM(Expectation-Maximization)-Clustering ······································· 415. 결과 ············································································································· 436. 결론 ············································································································· 50참고문헌 ·········································································································· 51ABSTRACT ······································································································ 55
최근 본 자료
댓글(0)
0