음성학적 지식 기반 변이음 모델을 이용한 가변 어휘 단어 인식기

김회린,이항섭,Kim, Hoi-Rin,Lee, Hang-Seop 한국음향학회 1997 韓國音響學會誌 Vol.16 No.2

본 논문에서는 훈련용 음성 데이터와 무관한 임의의 새로운 어휘를 인식해 낼 수 있는 가변 어휘 단어 인식기 개발에 대하여 기술한다. 가변 어휘 단어 인식기를 구현하기 위해서는, 인식 대상이 될 새로운 어휘를 즉시 발음 사전으로 변환시키는 on-line 발음 사전 생성기가 필요하고, 발음 사전 출력을 가지고 각 단어를 모델링할 수 있는 신뢰성 있는 음소 및 변이음 모델이 필요하다. 이와 같은 신뢰성 있는 음소 및 변이음 모델은 생성시키기 위하여 본 연구에서는, 각 음소의 전후 음소들의 음성학적 자질을 고려하여 3 음소열을 집단화(clustering)하여 변이음을 정의하고 이를 당 연구실이 보유하고 있는 POW(Phonetically Optimized Words) 3,848개 단어에 적용하여 1,548개의 변이음 모델을 생성시켰다. 이를 토대로 가변 어휘 단어 인식기를 구현하고 이를 POW 3,848 DB, PBW 445 DB 및 호텔 예약용 244 단어 DB 등에 적용하여 그 성능을 평가하였다. 평가 결과, POW DB에 대해서는 79.6%, PBW DB에 대해서는 445 단어 사전의 경우 79.4%, 100 단어 사전의 경우 88.9%의 성능을 보여 주었고, 호텔 예약 DB에 대해서는 71.4%의 성능을 보여 주었다. In this paper, we propose a variable vocabulary word recognizer that is able to recognize new words not exist in training data. For the variable vocabulary word recognizer, we must have an on-line lexicon generator to transform new candidate words to the corresponding pronunciation sequences of phones without any large lexicon table. And, we also must make outputs. In order to model the phones and allophones reliably, we define Korean allophones by triphone clustering based on phonetic knowledge of preceding and succeeding phones of each phone. Using the clustering method, we generated 1,548 allophones with POW (Phonetically Optimized Words) 3,848 word DB. We evaluated the proposed word recognizer with POW 3,848 DB, PBW (Phonetically Balanced Words) 445 DB, and 244 word DB in hotel reservation task. Experimental results showed word recognition accuracy of 79.6% for the POW DB corresponding to vocabulary-dependent case, 79.4% in case of 445 word lexicon and 88.9% in case of 100 word lexicon for the PBW DB, and 71.4% for the hotel reservation DB corresponding to vocabulary-independent case.

ICA 기반의 특징변환을 이용한 화자적응

박만수,김회린,Park ManSoo,Kim Hoi-Rin 대한음성학회 2002 말소리 Vol.43 No.-

The speaker adaptation technique is generally used to reduce the speaker difference in speech recognition. In this work, we focus on the features fitted to a linear regression-based speaker adaptation. These are obtained by feature transformation based on independent component analysis (ICA), and the transformation matrix is learned from a speaker independent training data. When the amount of data is small, however, it is necessary to adjust the ICA-based transformation matrix estimated from a new speaker utterance. To cope with this problem, we propose a smoothing method: through a linear interpolation between the speaker-independent (SI) feature transformation matrix and the speaker-dependent (SD) feature transformation matrix. We observed that the proposed technique is effective to adaptation performance.

핵심어 인식기에서 단어의 음소레벨 로그 우도 비율의 패턴을 이용한 발화검증 방법

김정현(Kim Chong-Hyon),권석봉(Kwon Suk-Bong),김회린(Kim Hoi-Rin) 한국음성학회 2009 말소리와 음성과학 Vol.1 No.1

Optimal Decision Tree를 이용한 Unseen Model 추정방법

김성탁,김회린,Kim Sungtak,Kim Hoi-Rin 대한음성학회 2003 말소리 Vol.45 No.-

Decision tree-based state tying has been proposed in recent years as the most popular approach for clustering the states of context-dependent hidden Markov model-based speech recognition. The aims of state tying is to reduce the number of free parameters and predict state probability distributions of unseen models. But, when doing state tying, the size of a decision tree is very important for word independent recognition. In this paper, we try to construct optimized decision tree based on the average of feature vectors in state pool and the number of seen modes. We observed that the proposed optimal decision tree is effective in predicting the state probability distribution of unseen models.

다양한 신뢰도 척도를 이용한 SVM 기반 발화검증 연구

권석봉,김회린,강점자,구명완,류창선,Kwon, Suk-Bong,Kim, Hoi-Rin,Kang, Jeom-Ja,Koo, Myong-Wan,Ryu, Chang-Sun 대한음성학회 2006 말소리 Vol.60 No.-

In this paper, we present several confidence measures (CM) for speech recognition systems to evaluate the reliability of recognition results. We propose heuristic CMs such as mean log-likelihood score, N-best word log-likelihood ratio, likelihood sequence fluctuation and likelihood ratio testing(LRT)-based CMs using several types of anti-models. Furthermore, we propose new algorithms to add weighting terms on phone-level log-likelihood ratio to merge word-level log-likelihood ratios. These weighting terms are computed from the distance between acoustic models and knowledge-based phoneme classifications. LRT-based CMs show better performance than heuristic CMs excessively, and LRT-based CMs using phonetic information show that the relative reduction in equal error rate ranges between $8{\sim}13%$ compared to the baseline LRT-based CMs. We use the support vector machine to fuse several CMs and improve the performance of utterance verification. From our experiments, we know that selection of CMs with low correlation is more effective than CMs with high correlation.

인식 단위로서의 한국어 음절에 대한 연구

김유진,김회린,정재호,Kim, Yu-Jin,Kim, Hoi-Rin,Chung, Jae-Ho 한국음향학회 1997 韓國音響學會誌 Vol.16 No.3

본 논문에서는 한국어 대용량 어휘 인식 시스템에 적합한 인식 단위에 대하여 연구 및 실험하였다. 특히 현재 인식 시스템의 인식 단위로 주로 사용되는 음소와 한국어의 특징을 잘 나타내는 음절을 선택하고, 인식 실험을 통해 음절이 한국어 인식 시스템의 인식 단위로서 적합한가를 음소와 비교하였다. 객관적인 비교 인식 실험 결과를 제시하기 위하여 동일한 남성 화자의 음성 데이터를 수집하고, 수작업 음소 경계 및 레이블링 과정을 거친 음성 데이터 베이스를 구축하였다. 또한 각 인식 단위에 동일한 HMM 기반의 훈련 및 인식 알고리즘을 적용하기 위해 Entropic사의 HTK (HMM Tool Kit) 2.0을 사용하였다. 각 인식 단위의 훈련을 위해 5상태 3출력, 8상태 6출력 HMM 모델의 연속 HMM (Continuous HMM)을 적용하였고, PBW 3회분, POW 1회분을 훈련에 사용하고 PBW 1회분을 각 인식 단위로서 인식하는 화자 종속 단어 인식 실험을 구성하였다. 실험 결과 8상태 6출력 모델을 사용한 경우 음소 단위는 95.65%, 음절 단위는 94.41%의 인식률을 나타내었다. 한편 인식 속도에서는 음절이 음소보다 약 25% 빠른 것으로 나타났다. In this paper, study and experiments are performed for finding recognition unit fit which can be used in large vocabulary recognition system. Specifically, a phoneme that is currently used as recognition unit and a syllable in which Korean is well characterized are selected. From comparisons of recognition experiments, the study is performed whether a syllable can be considered as recognition unit of Korean recognition system. For report of an objective result of the comparison experiment, we collected speech data of a male speaker and processed them by hand-segmentation for phoneme boundary and labeling to construct speech database. And for training and recognition based on HMM, we used HTK (HMM Tool Kit) 2.0 of commercial tool from Entropic Co. to experiment in same condition. We applied two HMM model topologies, 3 emitting state of 5 state and 6 emitting state of 8 state, in Continuous HMM on training of each recognition unit. We also used 3 sets of PBW (Phonetically Balanced Words) and 1 set of POW(Phonetically Optimized Words) for training and another 1 set of PBW for recognition, that is "Speaker Dependent Medium Vocabulary Size Recognition." Experiments result reports that recognition rate is 95.65% in phoneme unit, 94.41% in syllable unit and decoding time of recognition in syllable unit is faster by 25% than in phoneme.

한국어 음성인식 플랫폼의 설계

권오욱,김회린,유창동,김봉완,이용주,Kwon Oh-Wook,Kim Hoi-Rin,Yoo Changdong,Kim Bong-Wan,Lee Yong-Ju 대한음성학회 2004 말소리 Vol.51 No.-

For educational and research purposes, a Korean speech recognition platform is designed. It is based on an object-oriented architecture and can be easily modified so that researchers can readily evaluate the performance of a recognition algorithm of interest. This platform will save development time for many who are interested in speech recognition. The platform includes the following modules: Noise reduction, end-point detection, met-frequency cepstral coefficient (MFCC) and perceptually linear prediction (PLP)-based feature extraction, hidden Markov model (HMM)-based acoustic modeling, n-gram language modeling, n-best search, and Korean language processing. The decoder of the platform can handle both lexical search trees for large vocabulary speech recognition and finite-state networks for small-to-medium vocabulary speech recognition. It performs word-dependent n-best search algorithm with a bigram language model in the first forward search stage and then extracts a word lattice and restores each lattice path with a trigram language model in the second stage.

신뢰성 높은 서브밴드 특징벡터 선택을 이용한 잡음에 강인한 화자검증

김성탁,지미경,김회린,Kim, Sung-Tak,Ji, Mi-Kyong,Kim, Hoi-Rin 대한음성학회 2007 말소리 Vol.63 No.-

Recently, many techniques have been proposed to improve the noise robustness for speaker verification. In this paper, we consider the feature recombination technique in multi-band approach. In the conventional feature recombination for speaker verification, to compute the likelihoods of speaker models or universal background model, whole feature components are used. This computation method is not effective in a view point of multi-band approach. To deal with non-effectiveness of the conventional feature recombination technique, we introduce a subband likelihood computation, and propose a modified feature recombination using subband likelihoods. In decision step of speaker verification system in noise environments, a few very low likelihood scores of a speaker model or universal background model cause speaker verification system to make wrong decision. To overcome this problem, a reliable feature selection method is proposed. The low likelihood scores of unreliable feature are substituted by likelihood scores of the adaptive noise model. In here, this adaptive noise model is estimated by maximum a posteriori adaptation technique using noise features directly obtained from noisy test speech. The proposed method using subband-based reliable feature selection obtains better performance than conventional feature recombination system. The error reduction rate is more than 31 % compared with the feature recombination-based speaker verification system.

Music summarization using visual information of music and clustering method

김상호(Sangho Kim),지미경(Mi-kyong Ji),김회린(Hoi-Rin Kim) 한국HCI학회 2006 한국HCI학회 학술대회 Vol.2006 No.2

다중 레벨 양자화 기법을 적용한 오디오 핑거프린트 추출 방법

송원식,박만수,김회린,Song Won-Sik,Park Man-Soo,Kim Hoi-Rin 한국음향학회 2006 韓國音響學會誌 Vol.25 No.4

본 논문은 필립스의 음악 검색 기법을 기반으로 필터 뱅크 에너지 변화량과 음악의 통계적인 특성을 이용한 오디오 핑거프린트 추출 방법을 제안하였다. 기존의 필립스 방식은 제한된 주파수 영역을 너무 많은 필터 뱅크로 분할하여 분석함으로써 밴드들 사이에 연계성 및 왜곡에 대한 민감도가 증가하는 특징을 보일 수 있다. 제안된 방법은 필터 뱅크의 밴드 수를 줄여 왜곡에 대한 강인성을 증진시키고, 필터 뱅크 에너지의 변화량의 부호와 크기 정보를 통계적 특성을 고려한 양자화 기법을 이용해 2비트로 할당함으로써 오디오 핑거프린트의 고유성을 확보하였다. 추출된 2비트는 4개의 레벨로 정보를 표현함으로 각 레벨 사이에 연계성이 존재하게 된다. 이 같은 레벨 사이의 연계성은 유사도 측정 시 이용될 뿐만 아니라 오디오 핑거프린트를 기준으로 검색 영역을 확장하는 제안된 방식에서는 효율적인 검색 영역을 선택할 수 있는 정보로 활용 되었다. 제안된 방식은 다양한 주변 잡음환경 (거리, 백화점, 자동차, 사무실, 식당)에서의 실험을 통하여 주변 잡음에 강인한 특성을 보일 뿐만 아니라 검색 속도 또한 향상되는 특징을 보였다. In this paper, we proposed a new audio fingerprint extraction method, based on Philips' music retrieval algorithm, which uses the energy difference of neighboring filter-bank and probabilistic characteristics of music. Since Philips method uses too many filter-banks in limited frequency band, it may cause audio fingerprints to be highly sensitive to additive noises and to have too high correlation between neighboring bands. The proposed method improves robustness to noises by reducing the number of filter-banks while it maintains the discriminative power by representing the energy difference of bands with 2 bits where the quantization levels are determined by probabilistic characteristics. The correlation which exists among 4 different levels in 2 bits is not only utilized in similarity measurement. but also in efficient reduction of searching area. Experiments show that the proposed method is not only more robust to various environmental noises (street, department, car, office, and restaurant), but also takes less time for database search than Philips in the case where music is highly degraded.