• Journal of KIISE:Software and Applications
• /
• v.30 no.9
• /
• pp.867-875
• /
• 2003

This paper concerns continuous density HMM topology optimization. There have been several researches for HMM topology optimization. BIC (Bayesian Information Criterion) is one of the well known optimization criteria, which assumes statistically well behaved homogeneous model parameters. HMMs, however, are composed of several different kind of parameters to accommodate complex topology, thus BIC's assumption does not hold true for HMMs. Even though BIC reduced the total number of parameters of HMMs, it could not improve the recognition rates. In this paper, we proposed two new model selection criteria, HBIC (HMM-oriented BIC) and BIC_Anti. The former is proposed to improve BIC by estimating model priors separately. The latter is to combine BIC and anti-likelihood to accelerate discrimination power of HMMs. We performed some comparative research on couple of model selection criteria for online handwriting data recognition. We got better recognition results with fewer number of parameters.

• The Journal of the Acoustical Society of Korea
• /
• v.19 no.2
• /
• pp.90-94
• /
• 2000

In this paper, we propose an efficient method of estimating a critical threshold which is used to reject unreliable patterns in a HMM based recognition system. The rejection methods based on the anti-models which are formulated as the statistical hypothesis determine whether or not to accept an input pattern by comparing the likelihood ratio of HMM and anti-models to a critical threshold. It is quite difficult to fix a threshold for the probability of a HMM because the range of such probabilities varies severely depending on the chosen class model. We estimate the critical threshold, which is very class-dependent, using the likelihood scores for the training database. In our experiments, we applied the proposed estimating method of the threshold to the HMM based 3D hand gesture recognition system. We found that this method can be used successfully for rejecting unreliable input gestures regardless of the types of anti-models.

• Journal of the Korean Society for Aeronautical & Space Sciences
• /
• v.43 no.1
• /
• pp.23-32
• /
• 2015

Modeling of fighter pilots, which is a fundamental technology for war games using defense M&S (Modeling & Simulation) becomes one of the prominent research issues as the importance of defense M&S increases. Especially, the recent accumulation of combat logs makes it possible to adopt statistical learning methods to pilot modeling, and an HMM (Hidden Markov Model) which is able to utilize the sequential characteristic of combat logs is suitable for the modeling. However, since an HMM works only by using one type of features, discrete or continuous, to apply an HMM to heterogeneous features, type integration is required. Therefore, we propose a dPCA-HMM method, where dPCA (Discrete Principal Component Analysis) is combined with an HMM for the type integration. From experiments conducted on combat logs acquired from a simulator furnished by agency for defense development, the performance of the proposed model is evaluated and was satisfactory.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.23 no.1
• /
• pp.33-38
• /
• 2019

This paper proposes an effective recognition method of VPI patient's speech employing DNN-HMM-based speech recognition system, and evaluates the recognition performance compared to GMM-HMM-based system. The proposed method employs speaker adaptation technique to improve VPI speech recognition. This paper proposes to use simulated VPI speech for generating a prior model for speaker adaptation and selective learning of weight matrices of DNN, in order to effectively utilize the small size of VPI speech for model adaptation. We also apply Linear Input Network (LIN) based model adaptation technique for the DNN model. The proposed speaker adaptation method brings 2.35% improvement in average accuracy compared to GMM-HMM based ASR system. The experimental results demonstrate that the proposed DNN-HMM-based speech recognition system is effective for VPI speech with small-sized speech data, compared to conventional GMM-HMM system.

• 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.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.10 no.8
• /
• pp.901-906
• /
• 2015

We presented a method for Vad(Voice Activity Detection) using Bi-level HMM. Conventional methods need to do an additional post processing or set rule-based delayed frames. To cope with the problem, we applied to VAD a Bi-level HMM that has an inserted state layer into a typical HMM. And we used posterior ratio of voice states to detect voice period. Considering MFCCs(: Mel-Frequency Cepstral Coefficients) as observation vectors, we performed some experiments with voice data of different SNRs and achieved satisfactory results compared with well-known methods.

• Journal of KIISE
• /
• v.44 no.10
• /
• pp.1026-1033
• /
• 2017

In this paper a Bayesian modeling and duration-based prediction method is proposed for health clinic time series data using the Hierarchical Dirichlet Process Hidden Markov Model (HDP-HMM). HDP-HMM is a Bayesian extension of HMM which can find the optimal number of health states, a number which is highly uncertain and even difficult to estimate under the context of health dynamics. Test results of HDP-HMM using simulated data and real health clinic data have shown interesting modeling behaviors and promising prediction performance over the span of up to five years. The future of health change is uncertain and its prediction is inherently difficult, but experimental results on health clinic data suggests that practical long-term prediction is possible and can be made useful if we present multiple hypotheses given dynamic contexts as defined by HMM states.

• Journal of the Korean Institute of Telematics and Electronics C
• /
• v.36C no.6
• /
• pp.44-51
• /
• 1999

The HMM-Net is a neural network architecture that implements the computation of output probabilities of a hidden Markov model (HMM). The architecture is developed for the purpose of combining the discriminant power of neural networks with the time-domain modeling capability of HMMs. Criteria of maximum likehood(ML) and minimization of mean squared error(MMSE) are used for learning HMM-Net classifiers. The criterion MMSE is better than ML when initial learning condition is well established. However Ml is more useful one when the condition is incomplete[3]. Therefore we propose an efficient learning method of HMM-Net classifiers using a hybrid criterion(ML/MMSE). In the method, we begin a learning with ML in order to get a stable start-point. After then, we continue the learning with MMSE to search an optimal or near-optimal solution. Experimental results for the isolated numeric digits from /0/ to /9/, a training and testing time-series pattern set, show that the performance of the proposed method is better than the others in the respects of learning and recognition rates.

• The Journal of the Acoustical Society of Korea
• /
• v.14 no.5
• /
• pp.5-10
• /
• 1995

In this paper we present a clustering procedure based on the use of HMM in order to get multiple statistical models which can well absorb the variants of each speaker with different ways of saying words. The HMM-clustered models obtained from the developed technique are applied to the speaker independent isolated word recognition. The HMM clustering method splits off all observation sequences with poor likelihood scores which fall below threshold from the training set and create a new model out of the observation sequences in the new cluster. Clustering is iterated by classifying each observation sequence as belonging to the cluster whose model has the maximum likelihood score. If any clutter has changed from the previous iteration the model in that cluster is reestimated by using the Baum-Welch reestimation procedure. Therefore, this method is more efficient than the conventional template-based clustering technique due to the integration capability of the clustering procedure and the parameter estimation. Experimental data show that the HMM-based clustering procedure leads to $1.43\%$ performance improvements over the conventional template-based clustering method and $2.08\%$ improvements over the single HMM method for the case of recognition of the isolated korean digits.

• The Journal of the Acoustical Society of Korea
• /
• v.26 no.6
• /
• pp.259-268
• /
• 2007

This paper contributes to an improvement of the statistical bandwidth extension(BWE) system based on Hidden Markov Model(HMM). First, the existing HMM training method for BWE, which is suggested originally by Jax, is analyzed in comparison with the general Baum-Welch training method. Next, based on this analysis, a new HMM-based BWE method is suggested which adopts the Baum-Welch re-estimation algorithm instead of the Jax's to train HMM model. Conclusionally speaking, the Baum-Welch re-estimation algorithm is a generalized form of the Jax's training method. It is flexible and adaptive in modeling the statistical characteristic of training data. Therefore, it generates a better model to the training data, which results in an enhanced BWE system. According to experimental results, the new method performs much better than the Jax's BWE systemin all cases. Under the given test conditions, the RMS log spectral distortion(LSD) scores were improved ranged from 0.31dB to 0.8dB, and 0.52dB in average.