Monophone Distance를 이용한 효율적인 State Clustering

다국어 초록 (Multilingual Abstract)

In this thesis, efficient state clustering algorithms are described. The purpose of state clustering is to reduce the number of models and to use relatively reliable output probability distributions. There are two approaches: one is top down clustering, and the other is bottom up clustering. In earlier implementations of state clustering, bottom up state clustering is more used than top down state clustering. Bottom up state clustering is like follows. Similar acoustic states are merged first, and any cluster with insufficient training data is merged with its closest cluster. The advantage of this is that it can be merged more robust than top down state clustering. But it is fell performance because there are not the methods that are solved problem of unreliable and unseen triphone. An unreliable triphone is the one that appears in the training data too infrequently to train the model accurately and an unseen triphone is the one that don’t appear in the training data so can’t be modeled. In recent work, top down state clustering is more popular for state clustering. It classifies state of triphone by asking linguistic questions using decision tree that is a binary tree. The advantage of this is that it can be solving a problem of unreliable and unseen trihphone. But because triphones that is once split can’t be moved other cluster in the tree, it is a sub-optimal clustering algorithm.
So, we propose bottom up state clustering that can be solved a problem of unreliable and unseen triphone. The one of basic idea is to tie unreliable triphones before clustering. We use monophone distance to preprocess these unreliable triphones. It has been shown in a pilot experiment that the proposed method reduces the error rate significantly. The performances of the proposed algorithm recognition systems are evaluated on the TIMIT speech corpus.
The pure bottom up state clustering shows 6.3% recognition error rate, and the top down state clustering exhibits 6.9% recognition error rate. And the proposed algorithm displays 4.7% error rate. When unreliable triphone are merged first using monophone distance, we have found that the error rates can be reduced to 25.4% and 31.9%, respectably. These results show that our proposed algorithm has the better accuracy than the previous algorithms.

목차 (Table of Contents)

• 목 차
• 제 1 장 서론
• 제 2 장 관련 연구
• 2.1 Hidden Markov Models
• 2.1.1 구성 요소
• 2.1.2 확률값 계산
• 2.1.3 최적의 상태열
• 2.1.4 파라메터 학습
• 2.2 기존의 State Clustering 알고리즘
• 2.2.1 Bottom Up State clustering 알고리즘
• 2.2.2 Top Down State clustering 알고리즘
• 제 3 장 Undertrained triphone preprocessing
• 3.1 Monophone Distance를 이용한 Triphone Distance 표현
• 3.1.1 Bhattacharyya Distance
• 3.1.2 Triphone Distance
• 3.2 Undertrained triphone preprocessing
• 3.3 Unseen Triphone 해결 방법
• 제 4 장 실험 및 결과 분석
• 4.1 실험 환경
• 4.2 실험 결과
• 제 5 장 결론
• 참고문헌
• Abstract
• 표 목차
• [표 2-1] Clustering 알고리즘별 음성인식 에러율
• [표 3-1] Undertrained triphone preprocessing(UTP) 알고리즘
• [표 4-1] Seen triphone에 대한 인식률
• [표 4-2] 각 알고리즘별 에러를 일으키는 triphone
• [표 4-3] Unseen triphone을 포함한 인식실험
• 그림 목차
• [그림 2-1] 상태 천이 확률 분포 A와 심볼의 확률 분포 B
• [그림 2-2] Forward-backward procedure
• [그림 2-3] Triphone clustering 알고리즘
• [그림 2-4] Top down state clustering 알고리즘
• [그림 3-1] Monophone distance
• [그림 3-2] Unseen tirphone mapping 알고리즘