帶域 Energy-Band를 이용한 音聲區間의 檢出에 관한 硏究 = Endpoint-Detection of Speech Using Energy-Bands

다국어 초록 (Multilingual Abstract)

Conventional speech recognition system is trained and tested in noise-free environment. But in real life environment, recognition performance of the system will be dramatically degraded by many kinds of noise such as additive and convolutional noise etc. To be a real recognizer which will be operated well in noisy environment, the system must be robust to various noises added to input speech. One of the most important thing to accomplish for a robust recognizer to noise in real system is end-point detection of speech under various noise environments.
In this study, To realize a robust speech recognizer in such noise environments coarse and fine(C&F) search method which adopt high/low threshold based on SNR, SS(Spectral Subtraction), RASTA(RelAtive SpecTrAL), and CMN(Cepstral Mean Normalization) were adopted in the front-end process.
To test the effectiveness of the adopted methods three kinds of noises(such as white, pink, cafe) having different levels from 5dB to 40dB were added to input speech and scored the detection and recognition accuracies. The results were then compared with those from the conventional method using short-term energy and zero crossing(E&Z) rate.
As the results, end-point detection accuracies of input speech applied above three kinds of noises were increased by 10~20% with noise below 20dB and by 8~14% with office noise environment than E&Z.
When SS were added to C&F, it showed 5~10% of further improvement in accordance to SNR than those from E&Z method in recognition accuracy.
These results mean that C&F method which utilizes high/low threshold with SNR is effective under noisy environment.

목차 (Table of Contents)

• 목차 = i
• I. 서론 = 1
• II. 기존의 음성 구간 검출 방법 = 3
• 2.1 에너지를 이용한 음성구간 검출 = 3
• 2.2 에너지와 영교차율을 이용한 음성구간 검출 = 5
• III. 대역 에너지를 이용한 음성의 검출 = 10
• 3.1 청각 모델(Bark Scale)을 적용한 주파수 밴드 분할 = 10
• 3.2 신호대 잡음비의 추정 = 11
• 3.3 상위/하위 문턱치의 결정 = 12
• IV. Corarse 탐색과 Fine 탐색을 이용한 음성 구간 검출법 = 15
• 4.1 실시간 음성구간 검출 = 15
• 4.2 Coarse 탐색과 Fine 탐색 알고리즘 = 16
• V. 실험 및 결과 = 18
• 5.1 인식시스템 구성 = 18
• 5.1.1 전처리 = 18
• 5.1.2 이산 HMM 기반 인식 방법 = 21
• 5.1.3 인식시스템의 구성 = 31
• 5.2 음성구간 검출 실험 = 32
• 5.2.1 실험에 사용한 잡음의 특징 = 32
• 5.2.2 에너지와 영교차율을 이용한 음성구간 검출 = 33
• 5.2.3 Coarse 탐색과 Fine 탐색에 의한 음성구간 검출 = 34
• 5.2.4 음성구간 검출 방법에 따른 인식 비교 = 37
• 5.2 결과 분석 = 39
• VI. 결론 = 40
• 참고문헌 = 41
• 영문초록 = 43
• 부록A = 44