• The Journal of the Acoustical Society of Korea
• /
• v.18 no.5
• /
• pp.21-27
• /
• 1999

In this paper, we propose an improved MLP-based English pronunciation dictionary generator to apply to the variable vocabulary word recognizer. The variable vocabulary word recognizer can process any words specified in Korean word lexicon dynamically determined according to the current recognition task. To extend the ability of the system to task for English words, it is necessary to build a pronunciation dictionary generator to be able to process words not included in a predefined lexicon, such as proper nouns. In order to build the English pronunciation dictionary generator, we use context-dependent grapheme-to-phoneme multi-layer perceptron(MLP) architecture for each grapheme. To train each MLP, it is necessary to obtain grapheme-to-phoneme training data from general pronunciation dictionary. To automate the process, we use dynamic programming(DP) algorithm with some distance metrics. For training and testing the grapheme-to-phoneme MLPs, we use general English pronunciation dictionary with about 110 thousand words. With 26 MLPs each having 30 to 50 hidden nodes and the exception grapheme lexicon, we obtained the word accuracy of 72.8% for the 110 thousand words superior to rule-based method showing the word accuracy of 24.0%.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.1
• /
• pp.53-60
• /
• 1999

This paper is a study on the korean broadcasting speech recognition. Here we present the methods for the large vocabuary continuous speech recognition. Our main concerns are the language modeling and the search algorithm. The used acoustic model is the uni-phone semi-continuous hidden markov model and the used linguistic model is the N-gram model. The search algorithm consist of three phases in order to utilize all available acoustic and linguistic information. First, we use the forward Viterbi beam search to find word end frames and to estimate related scores. Second, we use the backword Viterbi beam search to find word begin frames and to estimate related scores. Finally, we use A/sup */ search to combine the above two results with the N-grams language model and to get recognition results. Using these methods maximum 96.0% word recognition rate and 99.2% syllable recognition rate are achieved for the speaker-independent continuous speech recognition problem with about 12,000 vocabulary size.

• The Journal of the Acoustical Society of Korea
• /
• v.9 no.4
• /
• pp.66-76
• /
• 1990

In this work, the effect of noise and distortion of a telephone channel on the speech recognition is studied, and methods to improve the recognition rate are proposed. Computer simulation is done using the 100-word test data whichwere made by pronouncing ten times 100-phonetically balanced Korean isolated words in a speaker dependent mode. First, a spectral subtraction method is suggested to improve the noisy speech recognition. Then, the effect of bandwidth limiting and channel distortion is studied. It has been found that bandwidth limiting and amplitude distortion lower the recognition rate significantly, but phase distortion affects little. To reduce the channel effect, we modify the reference pattern according to some training data. When both channel noise and distortion exist, the recognition rate without the proposed method is merely 7.7~26.4%, but the recognition rate with the proposed method is drastically increased to 76.2~92.3%.

• The Transactions of the Korea Information Processing Society
• /
• v.2 no.2
• /
• pp.215-228
• /
• 1995

Despite their importance in language, there have been relatively few computational studies in understanding words. This paper describes how neural networks can learn to perceive and produce words. Most traditional linguistic theories presuppose abstract underlying representations (UR) and a set of explicit rules to obtain the surface realization. There are, however, a number of questions that can be raised regarding this approach: (1) assumption of URs, (2) formation of rules, and (3) interaction of rules. In this paper, it is hypothesized that rules would emerge as the generalizations the network abstracts in the process of learning to associate forms with meanings of the words. Employing a simple recurrent network, a series of simulations on different types of morphophonemic processes was run. The results of the simulations show that this network is capable of learning to perceive whether words are in basic from or in inflected form, given only forms, and to produce words in the right form, given arbitrary meanings, this eliminating the need for presupposing abstract URs and rules.

• The Journal of the Acoustical Society of Korea
• /
• v.13 no.1
• /
• pp.24-31
• /
• 1994

In this paper, we report on the development of a speaker independent continuous speech recognition system using continuous hidden Markov models. The continuous hidden Markov model consists of mean and covariance matrices and directly models speech signal parameters, therefore does not have quantization error. Filter bank coefficients with their 1st and 2nd-order derivatives are used as feature vectors to represent the dynamic features of speech signal. We use the segmental K-means algorithm as a training algorithm and triphone as a recognition unit to alleviate performance degradation due to coarticulation problems critical in continuous speech recognition. Also, we use the one-pass search algorithm that Is advantageous in speeding-up the recognition time. Experimental results show that the system attains the recognition accuracy of $83\%$ without grammar and $94\%$ with finite state networks in speaker-indepdent speech recognition.

• Korean Journal of Cognitive Science
• /
• v.26 no.3
• /
• pp.263-278
• /
• 2015

Online Q&A for the National Institute of the Korean Language provides expert's answers for questions about the Korean language, in which many similar questions are repeatedly posted like other Q&A boards. So, if a system automatically finds questions that are similar to a user's question, it can immediately provide users with recommendable answers to their question and prevent experts from wasting time to answer to similar questions repeatedly. In this paper, we set 5 classes of questions based on its topic which are frequently asked, and propose to classify questions to those classes. Our system searches similar questions by combining topic similarity, vector similarity and sequence similarity. Experiment shows that our method improves search correctness with topic classification. In experiment, Mean Reciprocal Rank(MRR) of our system is 0.756, and precision for the first result is 68.31% and precision for top five results is 87.32%.

• Journal of KIISE:Software and Applications
• /
• v.27 no.3
• /
• pp.290-301
• /
• 2000

Most morphological analysis methods require repetitive procedures of input character code conversion, segmentation and lemmatization of constituent morphemes, filtering of candidate results through looking up lexicons, which causes run-time inefficiency. To alleviate such problem of run-time inefficiency, many systems have introduced the notion of 'pre-analysis' of words. However, this method based on pre-analysis dictionary of surface also has a critical drawback in its practical application because the size of the dictionaries increases indefinite to cover all words. This paper hybridizes both extreme approaches methodologically to overcome the problems of the two, and presents a method of morphological analysis based on pre-analysis of partial words. Under such hybridized scheme, most computational overheads, such as segmentation and lemmatization of morphemes, are shifted to building-up processes of the pre-analysis dictionaries and the run-time dictionary look-ups are greatly reduced, so as to enhance the run-time performance of the system. Moreover, additional computing overheads such as input character code conversion can also be avoided because this method relies upon no graphemic processing.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.3
• /
• pp.24-33
• /
• 2001

this paper describes a baseline for an implementation of a corpus-based Korean TTS system. The conventional TTS systems using small-sized speech still generate machine-like synthetic speech. To overcome this problem we introduce the corpus-based TTS system which enables to generate natural synthetic speech without prosodic modifications. The corpus should be composed of a natural prosody of source speech and multiple instances of synthesis units. To make a phone level synthesis unit, we train a speech recognizer with the target speech, and then perform an automatic phoneme segmentation. We also detect the fine pitch period using Laryngo graph signals, which is used for prosodic feature extraction. For break strength allocation, 4 levels of break indices are decided as pause length and also attached to phones to reflect prosodic variations in phrase boundaries. To predict the break strength on texts, we utilize the statistical information of POS (Part-of-Speech) sequences. The best triphone sequences are selected by Viterbi search considering the minimization of accumulative Euclidean distance of concatenating distortion. To get high quality synthesis speech applicable to commercial purpose, we introduce a domain specific database. By adding domain specific database to general domain database, we can greatly improve the quality of synthetic speech on specific domain. From the subjective evaluation, the new Korean corpus-based TTS system shows better naturalness than the conventional demisyllable-based one.

• Journal of Korea Society of Industrial Information Systems
• /
• v.6 no.3
• /
• pp.107-114
• /
• 2001

This paper describes Connected Digit Recognition with Considering Acoustic Feature in Korea. The recognition rate of connected digit is usually lower than word recognition. Therefore, speech feature parameter and acoustic feature are employed to make robust model for digit, and we could confirm the effect of Considering. Acoustic Feature throughout the experience of recognition. We used KLE 4 connected digit as database and 19 continuous distributed HMM as PLUs(Phoneme Like Units) using phonetical rules. For recognition experience, we have tested two cases. The first case, we used usual method like using Mel-Cepstrum and Regressive Coefficient for constructing phoneme model. The second case, we used expanded feature parameter and acoustic feature for constructing phoneme model. In both case, we employed OPDP(One Pass Dynamic Programming) and FSA(Finite State Automata) for recognition tests. When appling FSN for recognition, we applied various acoustic features. As the result, we could get 55.4% recognition rate for Mel-Cepstrum, and 67.4% for Mel-Cepstrum and Regressive Coefficient. Also, we could get 74.3% recognition rate for expanded feature parameter, and 75.4% for applying acoustic feature. Since, the case of applying acoustic feature got better result than former method, we could make certain that suggested method is effective for connected digit recognition in korean.

• Proceedings of the Acoustical Society of Korea Conference
• /
• spring
• /
• pp.74-77
• /
• 1999

본 논문에서는 본 연구실에서 개발한 주소 음성인식 시스템의 인식 속도를 개선시키기 위하예 새로운 가변 프루닝 문턱치를 적용하는 방법을 제안하고 실험을 통하여 그 유효성을 확인하였다. 기존의 가변 프루닝 문턱치는 일정 프레임이 경과하면 일정 값을 가진 문턱치를 계속하여 감소시켜나가는 방법을 반복하기 때문에, 불필요한 탐색공간을 탐색하게 된다. 본 논문에서 새로이 제안하는 가변 프루닝 문턱치를 채용하는 방법은 처음 일정 구간이 경과되면 일정 문턱치를 감소시키나, 다음 일정 프레임에서는 탐색되어야할 후보에 따라서 문턱치를 변화시켜 프루닝시키기 때문에 탐색공간을 효과적으로 감소시킬 수 있다. 제안된 방법의 유효성을 확인하기 위하여, 본 연구실에서 개발한 한국어 주소 입력 시스템에 적용하였다. 이 시스템은 48개의 연속 HMM 유사음소단위(Phoneme Like Units; PLUs)를 인식의 기본단위로 하고, .사용환경 변화에 의한 인식성능의 저하를 최소화하기 위해 최대사후 확률추정법(Maximum A Posteriori Probability Estimation; MAP)을 사용하며, 인식알고리즘으로는OPDP(One Pass Dynamic Programming)법을 이용하고 있다. 남성화자 3인에 의한 75개의 연결주소명을 이용하여 인식 실험을 수행한 결과 고정 프루닝 문턱치를 적용한 경우 인식률은 평균 $96.0\%$, 인식 시간은 5.26초였고, 기존의 가변 프루닝 문턱치의 경우 인식률은 평균 $96.0\%$, 인식 시간은 5.1초인 데 비하여, 새로운 가변 프루닝 문턱치를 적용찬 경우에는 인식률 저하없이 인식 시간이 4.34초로, 기존에 비해 각각 0.92초, 0.76초 인식 시간이 감소되어 제안한 방법의 유효성을 확인할 수 있었다.는 달리 각 산란 영역에서 그 지수는 1씩 작은 값을 갖는다.향에 따라 음장변화가 크게 다를 것이 예상되므로 이를 규명하기 위해서는 궁극적으로 3차원적인 음장분포 연구가 필요하다. 음향센서를 해저면에 매설할 경우 수충의 수온변화와 센서 주변의 수온변화 사이에는 어느 정도의 시간지연이 존재하게 되므로 이에 대한 영향을 규명하는 것도 센서의 성능예측을 위해서 필요하리라 사료된다.가지는 심부 가스의 개발 성공률을 증가시키기 위하여 심부 가스가 존재하는 지역의 지질학적 부존 환경 및 조성상의 특성과 생산시 소요되는 생산비용을 심도에 따라 분석하고 생산에 수반되는 기술적 문제점들을 정리하였으며 마지막으로 향후 요구되는 연구 분야들을 제시하였다. 또한 참고로 현재 심부 가스의 경우 미국이 연구 개발 측면에서 가장 활발한 활동을 전개하고 있으며 그 결과 다수의 신뢰성 있는 자료들을 확보하고 있으므로 본 논문은 USGS와 Gas Research Institute(GRI)에서 제시한 자료에 근거하였다.ऀĀ耀Ā삱?⨀؀Ā Ā?⨀ጀĀ耀Ā?돀ꢘ?⨀硩?⨀ႎ?⨀?⨀넆돐쁖잖⨀쁖잖⨀／ࠐ?⨀焆덐瀆倆Āⶇ퍟ⶇ퍟ĀĀĀĀ磀鲕좗?⨀肤?⨀⁅Ⴅ?⨀쀃잖⨀䣙熸ጁ↏?⨀