• The Journal of the Acoustical Society of Korea
• /
• v.41 no.3
• /
• pp.335-341
• /
• 2022

In this paper, we adopt a transformer that shows remarkable performance in natural language processing as an acoustic model of hybrid speech recognition. The transformer acoustic model uses attention structures to process sequential data and shows high performance with low computational cost. This paper proposes a method to improve the performance of transformer AM by applying each of the four algorithms of sequence discriminative training, a weighted finite-state transducer (wFST)-based learning used in the existing DNN-HMM model. In addition, compared to the Cross Entropy (CE) learning method, sequence discriminative method shows 5 % of the relative Word Error Rate (WER).

• Asia-pacific Journal of Multimedia Services Convergent with Art, Humanities, and Sociology
• /
• v.7 no.12
• /
• pp.819-829
• /
• 2017

In this paper, we describe the current state of speech recognition technology and identify the basic speech recognition technology and algorithms first, and then explain the code flow of API necessary for speech recognition technology. We use the application programming interface (API) of Google, Naver, and Daum KaKao, which have the most famous search engine among the speech recognition APIs, to create a voice recognition app in the Android studio tool. Then, we perform a speech recognition experiment on people's standard words and dialects according to gender, age, and region, and then organize the recognition rates into a table. Experiments were conducted on the Gyeongsang-do, Chungcheong-do, and Jeolla-do provinces where the degree of tongues was severe. And Comparative experiments were also conducted on standardized dialects. Based on the resultant sentences, the accuracy of the sentence is checked based on spacing of words, final consonant, postposition, and words and the number of each error is represented by a number. As a result, we aim to introduce the advantages of each API according to the speech recognition rate, and to establish a basic framework for the most efficient use.

• The Journal of the Acoustical Society of Korea
• /
• v.13 no.1E
• /
• pp.11-19
• /
• 1994

Hidden Markov Models (HMM's) with explicit state duration density (HMM/SD) can represent the time-varying characteristics of speech signals more accurately. However, such an advantage is reduced in relatively smooth state duration densities or ling bounded duration. To solve this problem, we propose HMM's with global path constraint (HMM/GPC) where the transition between states occur only within prescribed time slots. HMM/GPC explicitly limits state durations and accurately describes the temproal structure of speech simply and efficiently. HMM's formed by combining HMM/GPC with HMM/SD are also presented (HMM/SD+GPC) and performances are compared. HMM/GPC can be implemented with slight modifications to the conventional Viterbi algorithm. HMM/GPC and HMM/SD_GPC not only show superior performance than the conventional HMM and HMM/SD but also require much less computation. In the speaket independent isolated word recognition experiments, the minimum recognition eror rate of HMM/GPC(1.6%) is 1.1% lower than the conventional HMM's and the required computation decreased about 57%.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.39 no.3
• /
• pp.62-78
• /
• 2002

In this paper, we describe a system that can interact with agents in the virtual space incorporated in the system. This system is constructed by an analysis system for analyzing human gesture and an interact system for interacting with agents in the virtual space using analyzed information. An implemented analysis system for analyzing gesture extracts a head and hands region after taking image sequence of an operator's continuous behavior using CCD cameras. In interact system, we construct the virtual space that exist an avatar which incarnating operator himself, an autonomous object (like a Puppy), and non-autonomous objects which are table, door, window and object. Recognized gesture is transmitted to the avatar in the virtual space, then transit to next state based on state transition diagram. State transition diagram is represented in a graph in which each state represented as node and connect with link. In the virtual space, the agent link an avatar can open and close a window and a door, grab or move an object like a ball, order a puppy to do and respond to the Puppy's behavior as does the puppy.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.39 no.2
• /
• pp.116-127
• /
• 2002

A new approach to improve the speaker adaptation algorithm by means of the variable number of observation density functions for CDHMM speech recognizer has been proposed. The proposed method uses the observation density function with more than one mixture in each state to represent speech characteristics in detail. The number of mixtures in each state is determined by the number of frames and the determinant of the variance, respectively. The each MAP Parameter is extracted in every mixture determined by these two methods. In addition, the state segmentation method requiring speaker adaptation can segment the adapting speech more Precisely by using speaker-independent model trained from sufficient database as a priori knowledge. And the state duration distribution is used lot adapting the speech duration information owing to speaker's utterance habit and speed. The recognition rate of the proposed methods are significantly higher than that of the conventional method using one mixture in each state.

• Proceedings of the IEEK Conference
• /
• 2006.06a
• /
• pp.879-880
• /
• 2006

The subject of this paper is to recognize the stable state of EEG using wavelet transform and power spectrum analysis. An alpha wave, showed in stable state, is dominant wave for a human EEG and a beta wave displayed excited state. We decomposed EEG signal into an alpha wave and a beta wave in the process of wavelet transform. And we calculated each power spectrum of EEG signal, an alpha wave and a beta wave using Fast Fourier Transform. We recognized the stable state by making a comparison between power spectrum ratios respectively.

• International Journal of Contents
• /
• v.10 no.4
• /
• pp.11-17
• /
• 2014

The influence of acquisition environment on music score images captured by a camera has not yet been seriously examined. All existing Optical Music Recognition (OMR) systems attempt to recognize music score images captured by a scanner under ideal conditions. Therefore, when such systems process images under the influence of distortion, different viewpoints or suboptimal illumination effects, the performance, in terms of recognition accuracy and processing time, is unacceptable for deployment in practice. In this paper, a novel, lightweight but effective approach for dealing with the issues caused by camera based music scores is proposed. Based on the staff line information, musical rules, run length code, and projection, all regions of interest are determined. Templates created from inverse filter are then used to recognize the music symbols. Therefore, all fragmentation and deformation problems, as well as missed recognition, can be overcome using the developed method. The system was evaluated on a dataset consisting of real images captured by a smartphone. The achieved recognition rate and processing time were relatively competitive with state of the art works. In addition, the system was designed to be lightweight compared with the other approaches, which mostly adopted machine learning algorithms, to allow further deployment on portable devices with limited computing resources.

• Journal of Electrical Engineering and Technology
• /
• v.12 no.1
• /
• pp.363-372
• /
• 2017

This paper presents a novel face recognition algorithm based on the deep convolution neural network and key point detection jointed local binary pattern methodology to enhance the accuracy of face recognition. We firstly propose the modified face key feature point location detection method to enhance the traditional localization algorithm to better pre-process the original face images. We put forward the grey information and the color information with combination of a composite model of local information. Then, we optimize the multi-layer network structure deep learning algorithm using the Fisher criterion as reference to adjust the network structure more accurately. Furthermore, we modify the local binary pattern texture description operator and combine it with the neural network to overcome drawbacks that deep neural network could not learn to face image and the local characteristics. Simulation results demonstrate that the proposed algorithm obtains stronger robustness and feasibility compared with the other state-of-the-art algorithms. The proposed algorithm also provides the novel paradigm for the application of deep learning in the field of face recognition which sets the milestone for further research.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.23 no.3
• /
• pp.262-267
• /
• 2013

The recognition of human emotional state is one of the most important components for efficient human-human and human- computer interaction. In this paper, four emotions such as fear, disgust, joy, and neutral was a main problem of classifying emotion recognition and an approach of visual-stimuli for eliciting emotion based on physiological signals of skin conductance (SC), skin temperature (SKT), and blood volume pulse (BVP) was used to design the experiment. In order to reach the goal of solving this problem, half-against-half (HAH) multi-class support vector machine (SVM) with Gaussian radial basis function (RBF) kernel was proposed showing the effective techniques to improve the accuracy rate of emotion classification. The experimental results proved that the proposed was an efficient method for solving the emotion recognition problems with the accuracy rate of 90% of neutral, 86.67% of joy, 85% of disgust, and 80% of fear.

• Journal of Korea Multimedia Society
• /
• v.14 no.6
• /
• pp.711-718
• /
• 2011

The k-Nearest Neighbor classifier that does not require a training phase is appropriate for a variable number of classes problem like face recognition, Recently distance metric learning methods that is trained with a given data set have reported the significant improvement of the kNN classifier. However, the performance of a distance metric learning method is variable for each application, In this paper, we focus on the face recognition and compare the performance of the state-of-the-art distance metric learning methods, Our experimental results on the public face databases demonstrate that the Mahalanobis distance metric based on PCA is still competitive with respect to both performance and time complexity in face recognition.