• 전자공학회논문지B
• /
• 제32B권5호
• /
• pp.756-764
• /
• 1995

In this paper, we proposed MLP(MultiLayer Perceptron) based HMM's(Hidden Markov Models) with bounded state duration for isolated word recognition. The minimum and maximum state duration for each state of a HMM are estimated during the training phase and used as parameters of constraining state transition in a recognition phase. The procedure for estimating these parameters and the recognition algorithm using the proposed HMM's are also described. Speaker independent isolated word recognition experiments using a vocabulary of 10 city names and 11 digits indicate that recognition rate can be improved by adjusting the minimum state durations.

• The Journal of the Acoustical Society of Korea
• /
• 제17권2E호
• /
• pp.26-30
• /
• 1998

In this paper, we compose neural network predictive HMM(NNPHMM) to provide the dynamic feature of the speech pattern for the HMM. The NNPHMM is the hybrid network of neura network and the HMM. The NNPHMM trained to predict the future vector, varies each time. It is used instead of the mean vector in the HMM. In the experiment, we compared the recognition abilities of the one hundred Korean syllables according to the variation of hidden layer, state number and prediction orders of the NNPHMM. The hidden layer of NNPHMM increased from 10 dimensions to 30 dimensions, the state number increased from 4 to 6 and the prediction orders increased from 10 dimensions to 30 dimension, the state number increased from 4 to 6 and the prediction orders increased from the second oder to the fourth order. The NNPHMM in the experiment is composed of multi-layer perceptron with one hidden layer and CMHMM. As a result of the experiment, the case of prediction order is the second, the average recognition rate increased 3.5% when the state number is changed from 4 to 5. The case of prediction order is the third, the recognition rate increased 4.0%, and the case of prediction order is fourth, the recognition rate increased 3.2%. But the recognition rate decreased when the state number is changed from 5 to 6.

• 한국멀티미디어학회논문지
• /
• 제9권1호
• /
• pp.23-32
• /
• 2006

본 논문에서는 언어모델과 음향모델을 개선함으로써 단위 숫자음의 인식성능 최적화에 대해 설명한다. 언어모델은 한국어 단위 숫자음 문장의 문법적 특징을 분석하고, Finile State Network(FSN) 노드를 두 음절로 구성하여 오 인식률을 감소시켰다. 음향모델은 단 음절로 구성되어 발성기간이 짧고 조음이 많이 생기는 불명확한 음소, 음절의 분할로 인한 오 인식을 줄이기 위해 인식단위를 반음절 쌍으로 하였다. 인식단위의 특징을 효과적으로 모델링하기 위해 특징부분에서 K-means 알고리즘으로 군집화 하여, 상태를 분할하는 변형된 연쇄 상태 분할방법을 이용하였다. 실험 결과 제안된 언어모델의 적용 후 동일 문맥종속 음소모델에서 10.5%, 음향모델에서 인식단위를 반음절 쌍으로 하였을 경우 문맥종속 음소모델에 비해 12.5%, 변형된 연쇄 상태분할을 하였을 경우 1.5%의 인식률을 향상시킬 수 있었다.

• 한국음향학회지
• /
• 제16권3호
• /
• pp.64-72
• /
• 1997

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.

• 융합신호처리학회논문지
• /
• 제4권4호
• /
• pp.30-39
• /
• 2003

본 논문에서는 한국어 음성 데이터를 대상으로 HM-Net(Hidden Markov Network) 음성인식 시스템의 성능평가를 수행하였다. 음향모델 작성은 음성인식에서 널리 사용되고 있는 통계적인 모델링 방법인 HMM(Hidden Markov Model)을 개량한 HM-Net을 도입하였다. HM-Net은 기존의 SSS(Successive State Splitting) 알고리즘을 개량한 PDT(Phonetic Decision Tree)-SSS 알고리즘에 의해 문맥방향과 시간방향의 상태분할을 수행하여 생성되는데, 특히 문맥방향 상태분할의 경우 학습 음성데이터에 출현하지 않는 문맥정보를 효과적으로 표현하기 위해 음소결정트리를 채용하고 있으며, 시간방향 상태분할의 경우 학습 음성데이터에서 각 음소별 지속시간 정보를 효과적으로 표현하기 위한 상태분할을 수행하며, 마지막으로 파라미터의 공유를 통해 triphone 형태의 최적인 모델 네트워크를 작성하게 된다. 인식에 사용된 알고리즘은 음소 및 단어인식의 경우에는 One-Pass Viterbi 빔 탐색을 사용하며 트리 구조 형태의 사전과 phone/word-pair 문법을 채용하고 있다. 연속음성인식의 경우에는 단어 bigram과 단어 trigram 언어모델과 목구조 형태의 사전을 채용한 Multi-Pass 빔 탐색을 사용하고 있다. 전체적으로 본 논문에서는 다양한 조건에서 HM-Net 음성인식 시스템의 성능평가를 수행하였으며, 지금까지 소개된 음성인식 시스템과 비교하여 매우 우수한 인식성능을 보임을 실험을 통해 확인할 수 있었다.

• Bulletin of the Korean Chemical Society
• /
• 제28권5호
• /
• pp.805-810
• /
• 2007

Molecular recognition for a specific cation depending on the change of the oxidation state of the metal catalyst component contained in the hydrogel network has been studied in a self-oscillating hydrogel. The selfoscillating hydrogels are synthesized by the copolymerization of N-isopropylacrylamide (NIPAAm), lead methacrylic acid (Pb(MAA)2), and Ru(bpy)3 2+ monomer as a metal catalyst component. The recognition for a specific cation (in this study, Ca2+ has been used) is characterized by the adsorbed amount of Ca2+ into the gel. The recognition of the gels for Ca2+ is higher at the temperature below the LCST, and also higher at the oxidized state than at reduced state of the metal catalyst component which corresponds to a more swollen state. Moreover, a propagating wave induced by a periodic change of the oxidation state with the diffusion phenomena in the oscillating hydrogel shows a possibility for temporal and site-specific molecular recognition due to the local swelling of the gel.

• 대한전자공학회논문지
• /
• 제27권12호
• /
• pp.1933-1939
• /
• 1990

In this paper, we proposed an algorithm which used backtracking method to get time information, and it be modelled DMS (Dynamic Multisection) by feature vectors and time information whic are represented to similiar feature in word patterns spoken during continuous time domain, for Korean Speech recognition by independent speaker using DMS. Each state of model is represented time sequence, and have time information and feature vector. Typical feature vector is determined as the feature vector of each state to minimize the distance between word patterns. DDD Area names are selected as recognition wcabulary and 12th LPC cepstrum coefficients are used as the feature parameter. State of model is made 8 multisection and is used 0.2 as weight for time information. Through the experiment result, recognition rate by DMS model is 94.8%, and it is shown that this is better than recognition rate (89.3%) by MSVQ(Multisection Vector Quantization) method.

• 음성과학
• /
• 제8권3호
• /
• pp.123-132
• /
• 2001

In recent years research on HCI technology has been very active and speech recognition is being used as its typical method. Its recognition, however, is deteriorated with the increase of surrounding noise. To solve this problem, studies concerning the multimodal HCI are being briskly made. This paper describes automated lipreading for bimodal speech recognition on the basis of image- and speech information. It employs audio-visual DB containing 1,074 words from 70 voice and tri-viseme as a recognition unit, and state tied HMM as a recognition model. Performance of automated recognition of 22 to 1,000 words are evaluated to achieve word recognition of 60.5% in terms of 22word recognizer.

• Advances in Computational Design
• /
• 제9권3호
• /
• pp.187-211
• /
• 2024

WiFi sensing integration enables non-intrusive and is utilized in applications like Human Activity Recognition (HAR) to leverage Multiple Input Multiple Output (MIMO) systems and Channel State Information (CSI) data for accurate signal monitoring in different fields, such as smart environments. The complexity of extracting relevant features from CSI data poses computational bottlenecks, hindering real-time recognition and limiting deployment on resource-constrained devices. The existing methods sacrifice accuracy for computational efficiency or vice versa, compromising the reliability of activity recognition within pervasive environments. The lightweight Compact Convolutional Transformer (CCT) algorithm proposed in this work offers a solution by streamlining the process of leveraging CSI data for activity recognition in such complex data. By leveraging the strengths of both CNNs and transformer models, the CCT algorithm achieves state-of-the-art accuracy on various benchmarks, emphasizing its excellence over traditional algorithms. The model matches convolutional networks' computational efficiency with transformers' modeling capabilities. The evaluation process of the proposed model utilizes self-collected dataset for CSI WiFi signals with few daily activities. The results demonstrate the improvement achieved by using CCT in real-time activity recognition, as well as the ability to operate on devices and networks with limited computational resources.

• 대한음성학회:학술대회논문집
• /
• 대한음성학회 2002년도 11월 학술대회지
• /
• pp.197-200
• /
• 2002

In this paper, we propose a phoneme duration modeling in a speech recognition system based on disicion tree state tying. We assume that phone duration has a Gamma distribution. In a training mode, we model mean and variance of each state duration in context-independent phone model based on decision tree state tying. In a recognition mode, we get mean and variance of each context-dependent phone duration form state duration information obtaind during training mode. We make a comparative study of the proposed meth with conventinal methods. Our method results in good performance compared with conventional methods.