• The Journal of the Acoustical Society of Korea
• /
• v.38 no.5
• /
• pp.607-613
• /
• 2019

This paper proposes LF-MMI (Lattice Free Maximum Mutual Information)-based acoustic modeling using external knowledge for speech recognition. Note that an external knowledge refers to text data other than training data used in acoustic model. LF-MMI, objective function for optimization of training DNN (Deep Neural Network), has high performances in discriminative training. In LF-MMI, a phoneme probability as prior probability is used for predicting posterior probability of the DNN-based acoustic model. We propose using external knowledges for training the prior probability model to improve acoustic model based on DNN. It is measured to relative improvement 14 % as compared with the conventional LF-MMI-based model.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.32 no.6
• /
• pp.1081-1090
• /
• 2022

The face recognition model is used for identity recognition of smartphones, providing convenience to many users. As a result, the security review of the DNN model is becoming important, with adversarial attacks present as a well-known vulnerability of the DNN model. Adversarial attacks have evolved to decision-based attack techniques that use only the recognition results of deep learning models to perform attacks. However, existing decision-based attack technique[14] have a problem that requires a large number of queries when generating adversarial examples. In particular, it takes a large number of queries to approximate the gradient. Therefore, in this paper, we propose a method of generating adversarial examples using orthogonal space sampling and dimensionality reduction sampling to avoid wasting queries that are consumed to approximate the gradient of existing decision-based attack technique[14]. Experiments show that our method can reduce the perturbation size of adversarial examples by about 2.4 compared to existing attack technique[14] and increase the attack success rate by 14% compared to existing attack technique[14]. Experimental results demonstrate that the adversarial example generation method proposed in this paper has superior attack performance.

• Journal of Internet of Things and Convergence
• /
• v.9 no.2
• /
• pp.71-76
• /
• 2023

In this paper, we propose a smoothed received signal strength indicator (RSSI)-based distance estimation using deep neural network (DNN) for accurate distance estimation in an environment where a single receiver is used. The proposed scheme performs a data preprocessing consisting of data splitting, missing value imputation, and smoothing steps to improve distance estimation accuracy, thereby deriving the smoothed RSSI values. The derived smoothed RSSI values are used as input data of the Multi-Input Single-Output (MISO) DNN model, and are finally returned as an estimated distance in the output layer through input layer and hidden layer. To verify the superiority of the proposed scheme, we compared the performance of the proposed scheme with that of the linear regression-based distance estimation scheme. As a result, the proposed scheme showed 29.09% higher distance estimation accuracy than the linear regression-based distance estimation scheme.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.05a
• /
• pp.277-279
• /
• 2021

There is a problem that learning of a prediction model is not well performed depending on the characteristics of each particular matter concentration. To solve this problem, it is necessary to design a prediction model for low concentration and high concentration separately. Therefore, a classification model is needed to classify the concentration of particular matter into low and high concentrations. This paper proposes a classification model to classify low and high concentrations based on the concentration of particular matter. DNN was used as the classification model algorithm, and the classification model was designed by applying the optimal parameters after searching for hyper parameters. As for the result of evaluating the performance of the model, 97.54% of the low concentration classification was measured. And in the case of high concentration classification, 85.51% was measured.

• Journal of Digital Policy
• /
• v.2 no.4
• /
• pp.25-32
• /
• 2023

This paper is a study to develop a deep neural network (DNN) blood glucose prediction model based on heart rate (HR) and heart rate variability (HRV) data measured by PPG-based sensors. MLP deep learning consists of an input layer, a hidden layer, and an output layer with 11 independent variables. The learning results of the blood glucose prediction model are MAE=0.3781, MSE=0.8518, and RMSE=0.9229, and the coefficient of determination (R²) is 0.9994. The study was able to verify the feasibility of glycemic control using non-blood vital signs using PPG-based digital devices. In conclusion, a standardized method of acquiring and interpreting PPG-based vital signs, a large data set for deep learning, and a study to demonstrate the accuracy of the method may provide convenience and an alternative method for blood glucose management in dogs.

• Phonetics and Speech Sciences
• /
• v.10 no.3
• /
• pp.19-29
• /
• 2018

We propose to use independent component analysis (ICA) and deep neural network (DNN) to detect music sections in broadcast drama contents. Drama contents mainly comprise silence, noise, speech, music, and mixed (speech+music) sections. The silence section is detected by signal activity detection. To detect the music section, we train noise, speech, music, and mixed models with DNN. In computer experiments, we used the MUSAN corpus for training the acoustic model, and conducted an experiment using 3 hours' worth of Korean drama contents. As the mixed section includes music signals, it was regarded as a music section. The segmentation error rate (SER) of music section detection was observed to be 19.0%. In addition, when stereo mixed signals were separated into music signals using ICA, the SER was reduced to 11.8%.

• Phonetics and Speech Sciences
• /
• v.8 no.3
• /
• pp.31-38
• /
• 2016

Speech inversion (acoustic-to-articulatory mapping) is not a trivial problem, despite the importance, due to the highly non-linear and non-unique nature. This study aimed to investigate the performance of Deep Neural Network (DNN) compared to that of traditional Artificial Neural Network (ANN) to address the problem. The Wisconsin X-ray Microbeam Database was employed and the acoustic signal and articulatory pellet information were the input and output in the models. Results showed that the performance of ANN deteriorated as the number of hidden layers increased. In contrast, DNN showed lower and more stable RMS even up to 10 deep hidden layers, suggesting that DNN is capable of learning acoustic-articulatory inversion mapping more efficiently than ANN.

• Journal of Korea Multimedia Society
• /
• v.22 no.10
• /
• pp.1187-1198
• /
• 2019

In this study, we propose a new $PM_{10}$ forecasting model for Seoul region using DNN(Deep Neural Network) and secondary data. The previous numerical and Julian forecast model have been developed using primary data such as weather and air quality measurements. These models give excellent results for accuracy and false alarms, but POD is not good for the daily life usage. To solve this problem, we develop four secondary factors composed with primary data, which reflect the correlations between primary factors and high $PM_{10}$ concentrations. The proposed 4 models are A(Anomaly), BT(Back trajectory), CB(Contribution), CS(Cosine similarity), and ALL(model using all 4 secondary data). Among them, model ALL shows the best performance in all indicators, especially the PODs are improved.

• ETRI Journal
• /
• v.41 no.1
• /
• pp.109-116
• /
• 2019

This paper explores new techniques that are based on a hidden-layer linear transformation for fast speaker adaptation used in deep neural networks (DNNs). Conventional methods using affine transformations are ineffective because they require a relatively large number of parameters to perform. Meanwhile, methods that employ singular-value decomposition (SVD) are utilized because they are effective at reducing adaptive parameters. However, a matrix decomposition is computationally expensive when using online services. We propose the use of an extended diagonal linear transformation method to minimize adaptation parameters without SVD to increase the performance level for tasks that require smaller degrees of adaptation. In Korean large vocabulary continuous speech recognition (LVCSR) tasks, the proposed method shows significant improvements with error-reduction rates of 8.4% and 17.1% in five and 50 conversational sentence adaptations, respectively. Compared with the adaptation methods using SVD, there is an increased recognition performance with fewer parameters.

• Journal of Korea Multimedia Society
• /
• v.24 no.8
• /
• pp.1122-1137
• /
• 2021

PM_2.5 concentration in Seoul could be predicted by deep neural network model. In this paper, the contribution of input factors to the model's prediction results is analyzed using the LRP(Layer-wise Relevance Propagation) technique. LRP analysis is performed by dividing the input data by time and PM concentration, respectively. As a result of the analysis by time, the contribution of the measurement factors is high in the forecast for the day, and those of the forecast factors are high in the forecast for the tomorrow and the day after tomorrow. In the case of the PM concentration analysis, the contribution of the weather factors is high in the low-concentration pattern, and that of the air quality factors is high in the high-concentration pattern. In addition, the date and the temperature factors contribute significantly regardless of time and concentration.