Fast Speech Adaptation using Linear Spectral Transtorm for Additive and Convolutional Noises

국문 초록 (Abstract)

이 논문은 잡음 있는 환경에서 강인한 음성인식을 위해 변환 중심의 빠른 적응 알고리즘을 제안한다. 변환 중심의 빠른 적응이란, 적은 양의 데이터를 적응하였을 때 인식률을 높이기 위하여 Hidden Markov Models (HMMs)의 매개변수를 변환시키는 변환행렬을 이용하는 기술이다. 잡음 있는 환경에서 음성인식률을 높이기 위해서, additive 잡음과convolutional 잡음을 매개변수로 갖는 maximum likelihood spectral transform (MLST) 알고리즘이 제안되었다. 이 논문은 잡음에 직접적으로 영향 받는 선형 spectral 도메인에서 환경 잡음에 적응하기 위한 처리를 한다. 첫 번째 방법은 MLST의 평균값 변환 알고리즘으로 HMMs 매개변수 값을 조정하기 위해 연구되었다. 평균값 매개변수는 다양한 차원의 조건에서likelihood를 최대화하기 위해 변환된다. 두 번째 방법은 MLST 특성벡터 변환 알고리즘으로 입력되는 테스트 데이터를 조정하기위해 연구되었다. 이 방법들은 선형 spectral 도메인에서 작동하기 때문에 전통적인 적응 알고리즘에 비해서 적은 적응 데이터를 가지고도 빠른 적응을 할 수 있다.

다국어 초록 (Multilingual Abstract)

This thesis proposes a transform-based fast adaptation technique for robust speech recognition in unknown environments. Transform-based fast adaptation is the technique that uses transform matrix to alter the parameters of Hidden Markov Models (HMMs) to improve recognition accuracy with a small amount of adaptation data. To cope with noisy environment, maximum likelihood spectral transform (MLST) algorithm with additive and convolutional noise parameters is used.
Adaptation technique has the advantage of using the former acoustic model and avoiding retraining cost, and it is needed for achieving high performance in real environments. Previously many adaptation algorithms have been proposed in the cepstral domain. Though the cepstral domain may be appropriate for the speech recognition, it is difficult to handle environmental noise directly in the cepstral domain. The clean speech directly combined with environmental noise in the linear spectral domain, which may consists of additive noise, such as background sound, or convolutional noise, such as channel differences between training and testing. Conventional adaptation techniques used probabilistic approach or preprocessing of the speech signal, such as noise filtering to handle this speech in real environment.
Therefore this thesis deals with such environmental noise in the linear spectral domain in which speech is directly affected by the noise. The first, mean transform MLST is studied for compensating the HMMs. The mean parameters are transformed forward maximizing likelihood with variable conditions of dimensionality. The second, feature transform MLST is studied for compensating the testing data. These approaches work in linear spectral domain so that smaller number of transformation parameters can be used for fast adaptation than those of conventional algorithms.
The proposed algorithms are evaluated on the TIMIT and FFMTIMIT corpus.
TIMIT is used to make training models, and FFMTIMIT is used for adaptation
and testing data. The MLST algorithm shows promising result with only a small
number of adaptation data.

목차 (Table of Contents)

• Abstract = ⅰ
• 초록 = ⅰ
• Contents = ⅳ
• 1. Introduction = 1
• 1.1 Speech Recognition in Real Environments = 1
• 1.2 Goal of the Thesis = 3
• 1.3 Organization of the Thesis = 4
• 2. Techniques for Environment Adaptation = 5
• 2.1 Hidden Markov Models = 5
• 2.1.1 HMM Assumption = 6
• 2.1.2 Transition and Output Probability = 6
• 2.1.3 Pattern Matching = 7
• 2.1.4 Parameter Estimation = 9
• 2.1.5 Recognition = 12
• 2.2 Maximum a Posterior Based Method = 13
• 2.3 Maximum Likelihood Linear Regression = 14
• 2.4 Parallel Model Combination = 16
• 2.5 Maximum Likelihood Linear Transform = 19
• 3. Maximum Likelihood Spectral Transform = 20
• 3.1 Motivation = 20
• 3.2 Linear Spectral Transform = 21
• 3.3 Estimation of MLST Mean Adaptation Parameters = 24
• 3.4 Estimation of MLST Feature Adaptation Parameters = 26
• 3.5 Summary = 29
• 4. Evaluation of MLST Adaptation = 30
• 4.1 Methods of Evaluation = 30
• 4.2 Dimensionality of Transform Parameters = 31
• 4.3 Word Adaptation using Mean Transform = 33
• 4.4 Word Adaptation using Feature Transform = 34
• 4.5 Summary = 36
• 5. Conclusion and Future Works = 37
• References = 39