• Speech Sciences
• /
• v.11 no.4
• /
• pp.43-52
• /
• 2004

Model parameters in HMM based speech recognition systems are normally estimated using Maximum Likelihood Estimation(MLE). The MLE method is based mainly on the principle of statistical data fitting in terms of increasing the HMM likelihood. The optimality of this training criterion is conditioned on the availability of infinite amount of training data and the correct choice of model. However, in practice, neither of these conditions is satisfied. In this paper, we propose a training algorithm, MCE(Minimum Classification Error), to improve the performance of a speech recognizer detecting mispronunciation of a foreign language. During the conventional MLE(Maximum Likelihood Estimation) training, the model parameters are adjusted to increase the likelihood of the word strings corresponding to the training utterances without taking account of the probability of other possible word strings. In contrast to MLE, the MCE training scheme takes account of possible competing word hypotheses and tries to reduce the probability of incorrect hypotheses. The discriminant training method using MCE shows better recognition results than the MLE method does.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.31B no.6
• /
• pp.17-24
• /
• 1994

This paper is a study on Korean word recognition and suggest the method that stabilizes the state-transition in the HMM by applying the `inertia' to the feature vector sequences. In order to reduce the quantized distortion considering probability distribution of input vectors, we used SOFM, an unsupervised learning method, as a vector quantizer, By applying inertia to the feature vector sequences, the overlapping of probability distributions for the response path of each word on the self organizing feature map can be reduced and the state-transition in the Hmm can be Stabilized. In order to evaluate the performance of the method, we carried out experiments for 50 DDD area names. The results showed that applying inertia to the feature vector sequence improved the recognition rate by 7.4% and can make more HMMs available without reducing the recognition rate for the SOFM having the fixed number of neuron.

• Phonetics and Speech Sciences
• /
• v.15 no.1
• /
• pp.1-7
• /
• 2023

This study investigated the influences of probabilistic accessibility of the word being produced - as determined by its usage frequency and neighborhood density - on native and high-proficiency L2 speakers' realization of six English monophthong vowels. The native group hyperarticulated the vowels over an expanded acoustic space when the vowel occurred in words with low frequency and high density, supporting the claim that vowel forms are modified in accordance with the probabilistic accessibility of words. However, temporal expansion occurred in words with greater accessibility (i.e., with high frequency and low density) as an effect of low phonotactic probability in low-density words, particularly in attended speech. This suggests that temporal modification in the opposite direction may be part of the phonetic characteristics that are enhanced in communicatively driven focus realization. Conversely, none of these spectral and temporal patterns were found in the L2 group, thereby indicating that even the high-proficiency L2 speakers may not have developed experience-based sensitivity to the modulation of sub-categorical phonetic details indexed with word-level probabilistic information. The results are discussed with respect to how phonological representations are shaped in a word-specific manner for the sake of communicatively driven lexical intelligibility, and what factors may contribute to the lack of native-like sensitivity in L2 speech.

• The KIPS Transactions:PartB
• /
• v.10B no.7
• /
• pp.829-836
• /
• 2003

In this paper, we propose a stochastic model for sentence speech understanding using dictionary and thesaurus. The proposed model extracts words from an input speech or text into a sentence. A computer is sellected category of dictionary database compared the word extracting from the input sentence calculating a probability value to the compare results from stochastic model. At this time, computer read out upper dictionary information from the upper dictionary searching and extracting word compared input sentence caluclating value to the compare results from stochastic model. We compare adding the first and second probability value from the dictionary searching and the upper dictionary searching with threshold probability that we measure the sentence understanding rate. We evaluated the performance of the sentence speech understanding system by applying twenty questions game. As the experiment results, we got sentence speech understanding accuracy of 79.8%. In this case, probability ($\alpha$) of high level word is 0.9 and threshold probability ($\beta$) is 0.38.

• The Journal of the Acoustical Society of Korea
• /
• v.10 no.2
• /
• pp.29-35
• /
• 1991

In this paper, we discuss the method of recognizing allophone unit isolated words using hidden Markov model(HMM). Frist we constructed allophone lexicon by extracting allophones from training data and by training allophone HMMs. And then to recognize isolated words using allophone HMMs, it is necessary to construct word dictionary which contains information of allophone sequence and inter-allophone transition probability. Allophone sequences are represented by allophone HMMs. To see the effects of inter-allophone transition probability and to determine optimal probabilities, we performend some experiments. And we showed that small number of traing data and simple train procedure is needed to train word HMMs of allophone sequences and that not less performance than word unit HMM is obtained.

• Journal of the Institute of Electronics Engineers of Korea SD
• /
• v.40 no.4
• /
• pp.74-84
• /
• 2003

Conventional test algorithms do not effectively detect faults by word-line and bit-line coupling noise resulting from the increase of the density of memories. In this paper, the possibility of faults caused by word-line coupling noise is shown, and new fault model, WLSFs(Word-Line Sensitive Fault) is proposed. We also introduce the algorithm considering both word-line and bit-line coupling noise simultaneously. The algorithm increases probability of faults which means improved fault coverage and more effective test algorithm, compared to conventional ones. The proposed algorithm can also cover conventional basic faults which are stuck-at faults, transition faults and coupling faults within a five-cell physical neighborhood.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.73-75
• /
• 2022

Scene Text Recognition is a technology used in the field of artificial intelligence that requires manless robot, automatic vehicles and human-computer interaction. Though scene text images are distorted by noise interference, such as illumination, low resolution and blurring. Unlike previous studies that recognized only English, this paper shows a strong recognition accuracy including various characters, English, Korean, special character and numbers. Instead of selecting only one class having the highest probability value, a candidate word can be generated by considering the probability value of the second rank as well, thus a method can be corrected an existing language misrecognition problem.

• Language and Information
• /
• v.10 no.1
• /
• pp.21-46
• /
• 2006

This study proposes a Word Sense Disambiguation (WSD) algorithm, based on concept learning with special emphasis on statistically meaningful lowest frequency words. Previous works on WSD typically make use of frequency of collocation and its probability. Such probability based WSD approaches tend to ignore the lowest frequency words which could be meaningful in the context. In this paper, we show an algorithm to extract and make use of the meaningful lowest frequency words in WSD. Learning method is adopted from the Find-Specific algorithm of Mitchell (1997), according to which the search proceeds from the specific predefined hypothetical spaces to the general ones. In our model, this algorithm is used to find contexts with the most specific classifiers and then moves to the more general ones. We build up small seed data and apply those data to the relatively large test data. Following the algorithm in Yarowsky (1995), the classified test data are exhaustively included in the seed data, thus expanding the seed data. However, this might result in lots of noise in the seed data. Thus we introduce the 'maximum a posterior hypothesis' based on the Bayes' assumption to validate the noise status of the new seed data. We use the Naive Bayes Classifier and prove that the application of Find-Specific algorithm enhances the correctness of WSD.

• KIPS Transactions on Software and Data Engineering
• /
• v.3 no.10
• /
• pp.407-420
• /
• 2014

In general, the Korean morpheme analysis procedure is divided into two steps. In the first step as an ambiguity generation step, an Eojeol is analyzed into many morpheme sequences as candidates. In the second step, one appropriate candidate is chosen by using contextual information. Hidden Markov Model(HMM) is typically applied in the second step. This paper proposes Sub-word Conditional Probability(SCP) model as an alternate algorithm. SCP uses sub-word information of adjacent eojeol first. If it failed, then SCP use morpheme information restrictively. In the accuracy and speed comparative test, HMM's accuracy is 96.49% and SCP's accuracy is just 0.07% lower. But SCP reduced processing time 53%.

• Journal of the Korean Society of Clothing and Textiles
• /
• v.46 no.6
• /
• pp.1004-1022
• /
• 2022

This study seeks to investigate K-fashion's external image by examining the trends in global media reporting. It applies Dynamic Topic Modeling (DTM), which captures the evolution of topics in a sequentially organized corpus of documents, and consists of text preprocessing, the determination of the number of topics, and a timeseries analysis of the probability distribution of words within topics. The data set comprised 551 online media articles on 'Korean fashion' or 'K-fashion' published on Google News between 2010 and 2021. The analysis identifies seven topics: 'brand look and style,' 'lifestyle,' 'traditional style,' 'Seoul Fashion Week (SFW) event,' 'model size,' 'K-pop,' and 'fashion market,' as well as annual topic proportion trends. It also explores annual word changes within the topic and indicates increasing and decreasing word patterns. In most topics, the probability distribution of the word 'brand' is confirmed to be on the increase, while 'digital,' 'platform,' and 'virtual' have been newly created in the 'SFW event' topic. Moreover, this study confirms the transition of each K-fashion topic over the past 12 years, along with various factors related to Hallyu content, traditional culture, government support, and digital technology innovation.