• Journal of Korea Multimedia Society
• /
• v.11 no.6
• /
• pp.807-815
• /
• 2008

A simple efficient method of spotting and recognizing hand gestures in video is presented using a network of hidden Markov models and dynamic programming search algorithm. The description starts from designing a set of isolated trajectory models which are stochastic and robust enough to characterize highly variable patterns like human motion, handwriting, and speech. Those models are interconnected to form a single big network termed a spotting network or a spotter that models a continuous stream of gestures and non-gestures as well. The inference over the model is based on dynamic programming. The proposed model is highly efficient and can readily be extended to a variety of recurrent pattern recognition tasks. The test result without any engineering has shown the potential for practical application. At the end of the paper we add some related experimental result that has been obtained using a different model - dynamic Bayesian network - which is also a type of stochastic model.

• Proceedings of the Acoustical Society of Korea Conference
• /
• autumn
• /
• pp.109-112
• /
• 1999

본 연구는 음소 단위의 CHMM(Continuous Hidden Markov Model)을 이용한 한국어 연속 음성인식에 관한 내용이다. 연구실 환경에서 음성으로 전화를 걸기 위하여 연속 숫자음 인식을 수행하였다. ETRI 445 데이터를 사용하여 초기의 모델은 ML(Maximum Likelihood) 추정법을 이용하여 작성하였고 적응화를 위해 최대 사후 확률 추정법을 사용하였다. 연속 숫자음의 인식을 위하여 한국어 숫자음 음성의 음향학적 특성을 고려하여 발성 사전을 작성하였고, 음절 단위로 되어있는 한국어 숫자음의 모든 경우를 고려하여 복수개의 단어를 사전에 등록하였다. 또한 숫자음의 알 뒤 연음현상을 고려하여 작성한 21 종류의 7자리 숫자음과 이를 음절 단위로 세그먼트한 숫자음을 DB로 사용하여 적응화를 수행하였다. 이의 효율성을 입증하기 위하여 ETRI에서 작성한 35종류의 4연속 숫자음 목록을 대상으로 인식실험을 수행하였다.

• Proceedings of the Korea Multimedia Society Conference
• /
• 1998.10a
• /
• pp.221-225
• /
• 1998

본 논문에서는 연속분포(Continuous) HMM(hidden Markov model)을 기반으로 하여 한국어 고립단어인식 시스템을 설계, 구현하였다. 시스템의 학습과 평가를 위해 자동차 항법용 음성 명령어 도메인에서 추출한 10개의 고립단어를 대상으로 음성 데이터 베이스를 구축하였다. 음성 특징 파라미터로는 MFCCs(Mel Frequency Cepstral Coefficients)와 차분(delta) MFCC 그리고 에너지(energy)를 사용하였다. 학습 데이터로부터 추출한 18개의 유사 음소(phoneme-like unit : PLU)를 인식단위로 HMM 모델을 만들었고 조음 결합 현상(채-articulation)을 모델링 하기 위해 트라이폰(triphone) 모델로 확장하였다. 인식기 평가는 학습에 참여한 음성 데이터와 학습에 참여하지 않은 화자가 발성한 음성 데이터를 이용해 수행하였으며 평균적으로 97.5%의 인식성능을 얻었다.

• Annual Conference on Human and Language Technology
• /
• 1993.10a
• /
• pp.101-110
• /
• 1993

본 논문에서는 연속분포 hidden Markov 모델을 이용한 화자독립 연속 음성 인식 시스템에 관해 기술한다. 연속분포 모델은 평균과 분산 벡터로 구성되며 음성신호를 직접 모델링하여 양자화 왜곡이 없어진다. 특징벡터는 filter bank 계수 및 그 1, 2차 미분계수를 사용하여 음성신호의 동적 특성을 반영하였다. Segmental K-means 알고리즘을 이용하여 학습하였으며, 연속어 인식에서 가장 문제가 되는 조음화 현상으로 인한 인식률 저하를 막기 위해 앞뒤의 음소를 고려해 주는 triphone을 인식단위로 사용하였다. Search 알고리즘으로는 시간 면에서 효율이 좋은 one-pass search 알고리즘을 사용하였다. 성능 평가를 위한 화자 독립 인식 실험에서 문법이 없을 경우 83%, finite state network율 적용한 경우에는 94%의 인식률을 나타내었다.

• The Transactions of the Korea Information Processing Society
• /
• v.6 no.1
• /
• pp.247-252
• /
• 1999

In order to realize a speech coding at low bit rates, a pitch information is useful parameter. In case of extracting an average pitch information form continuous speech, the several pitch errors appear in a frame which consonant and vowel are coexistent; in the boundary between adjoining frames and beginning or ending of a sentence. In this paper, I propose an Individual Pitch (IP) extraction method using residual signals of the FIR-STREAK digital filter in order to restrict the pitch extraction errors. This method is based on not averaging pitch intervals in order to accomodate the changes in each pitch interval. As a result, in case of Ip extraction method suing FIR-STREAK digital filter, I can't find the pitch errors in a frame which consonant and vowel are consistent; in the boundary between adjoining frames and beginning or ending of a sentence. This method has the capability of being applied to many fields, such as speech coding, speech analysis, speech synthesis and speech recognition.

• Journal of Korea Multimedia Society
• /
• v.4 no.4
• /
• pp.306-315
• /
• 2001

In noise reduction method from noisy speech for speech recognition in noisy environments, conventional spectral subtraction method has a disadvantage which distinction of noise and speech is difficult, and characteristic of noise can't be estimated accurately. Also, noise reduction method in the wavelet transform domain has a disadvantage which loss of signal is generated in the high frequency domain. In order to compensate theme disadvantage, this paper propose spectral subtraction method in continuous wavelet transform domain which speech and non- speech intervals is distinguished by standard deviation of wavelet coefficient, and signal is divided three scales at different scale. The proposed method extract accurately characteristic of noise in order to apply spectral subtraction method by end detection and band division. The proposed method shows better performance than noise reduction method using conventional spectral subtraction and wavelet transform from viewpoint signal to noise ratio and Itakura-Saito distance by experimental.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.21 no.2
• /
• pp.534-540
• /
• 1996

This paper presents an Adaptive Time-Delay Recurrent Neural Network (ATRN) for learning and recognition of temporal correlations of temporal patterns. The ATRN employs adaptive time-delays and recurrent connections, which are inspired from neurobiology. In the ATRN, the adaptive time-delays make the ATRN choose the optimal values of time-delays for the temporal location of the important information in the input parrerns, and the recurrent connections enable the network to encode and integrate temporal information of sequences which have arbitrary interval time and arbitrary length of temporal context. The ATRN described in this paper, ATNN proposed by Lin, and TDNN introduced by Waibel were simulated and applied to the chaotic time series preditcion of Mackey-Glass delay-differential equation. The simulation results show that the normalized mean square error (NMSE) of ATRN is 0.0026, while the NMSE values of ATNN and TDNN are 0.014, 0.0117, respectively, and in temporal learning, employing recurrent links in the network is more effective than putting multiple time-delays into the neurons. The best performance is attained bythe ATRN. This ATRN will be sell applicable for temporally continuous domains, such as speech recognition, moving object recognition, motor control, and time-series prediction.

• Journal of the Institute of Convergence Signal Processing
• /
• v.2 no.4
• /
• pp.9-15
• /
• 2001

In this paper we investigate context dependent acoustic models to improve the performance of the Korean speech recognition . The algorithm are using the Korean phonological rules and decision tree, By Successive State Splitting(SSS) algorithm the Hidden Merkov Netwwork(HM-Net) which is an efficient representation of phoneme-context-dependent HMMs, can be generated automatically SSS is powerful technique to design topologies of tied-state HMMs but it doesn't treat unknown contexts in the training phoneme contexts environment adequately In addition it has some problem in the procedure of the contextual domain. In this paper we adopt a new state-clustering algorithm of SSS, called Phonetic Decision Tree-based SSS (PDT-SSS) which includes contexts splits based on the Korean phonological rules. This method combines advantages of both the decision tree clustering and SSS, and can generated highly accurate HM-Net that can express any contexts To verify the effectiveness of the adopted methods. the experiments are carried out using KLE 452 word database and YNU 200 sentence database. Through the Korean phoneme word and sentence recognition experiments. we proved that the new state-clustering algorithm produce better phoneme, word and continuous speech recognition accuracy than the conventional HMMs.

• Proceedings of the KSPS conference
• /
• 2006.11a
• /
• pp.80-83
• /
• 2006

The purpose of this paper is to build a pronunciation lexicon with estimated likelihoods of the phonetic rules based on the phonetic realizations and therefore to improve the performance of CSR using the dictionary. In the baseline system, the phonetic rules and their application probabilities are defined with the knowledge of Korean phonology and experimental tuning. The advantage of this approach is to implement the phonetic rules easily and to get stable results on general domains. However, a possible drawback of this method is that it is hard to reflect characteristics of the phonetic realizations on a specific domain. In order to make the system reflect phonetic realizations, the likelihood of phonetic rules is reestimated based on the statistics of the realized phonemes using a forced-alignment method. In our experiment, we generates new lexica which include pronunciation variants created by reestimated phonetic rules and its performance is tested with 12 Gaussian mixture HMMs and back-off bigrams. The proposed method reduced the WER by 0.42%.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.1
• /
• pp.75-81
• /
• 2004

Gaussian selection (GS) is a popular approach in the continuous density hidden Markov model for fast decoding. It enables fast likelihood computation by reducing the number of Gaussian components calculated. In this paper, we propose a new GS method for the phonetic tied-mixture (PTM) hidden Markov models. The PTM model can represent each state of the same topological location with a shared set of Gaussian mixture components and contort dependent weights. Thus the proposed method imposes constraint on the weights as well as the number of Gaussian components to reduce the computational load. Experimental results show that the proposed method reduces the percentage of Gaussian computation to 16.41%, compared with 20-30% for the conventional GS methods, with little degradation in recognition.